diff --git a/tools/python/transform/onnx_converter.py b/tools/python/transform/onnx_converter.py
index c7d11ab72b2a5280a14c28bca9c0bce7c8cb8b1c..c1a7b0d680beb6836e70f5d20fe0cd79a089dbed 100644
--- a/tools/python/transform/onnx_converter.py
+++ b/tools/python/transform/onnx_converter.py
@@ -191,9 +191,9 @@ OnnxOpType = Enum('OnnxOpType',
 onnx_attr_translator = {
     "axis": lambda x: int(x),
     "axes": lambda x: [int(a) for a in x],
-    "dtype": lambda x: data_type.onnx2tf(x),
+    "dtype": lambda x: onnx_dtype(x),
     "keepdims": lambda x: bool(x),
-    "to": lambda x: data_type.onnx2tf(x),
+    "to": lambda x: onnx_dtype(x),
 }
 
 
@@ -567,11 +567,7 @@ class OnnxConverter(base_converter.ConverterInterface):
                     tensor.data_type = mace_pb2.DT_FLOAT
                     tensor.float_data.extend(
                         onnx_tensor.astype(np.float32).flat)
-                elif data_type == np.int32:
-                    tensor.data_type = mace_pb2.DT_INT32
-                    tensor.int32_data.extend(
-                        onnx_tensor.astype(np.int32).flat)
-                elif data_type == np.int64:
+                elif data_type == np.int64 or data_type == np.int32:
                     tensor.data_type = mace_pb2.DT_INT32
                     tensor.int32_data.extend(
                         onnx_tensor.astype(np.int32).flat)
@@ -668,9 +664,9 @@ class OnnxConverter(base_converter.ConverterInterface):
 
         if 'to' in node.attrs:
             dtype = node.attrs['to']
-            if dtype == TensorProto.FLOAT:
+            if dtype == np.float32 or dtype == np.float64:
                 op.output_type.extend([self._option.data_type])
-            elif dtype == TensorProto.INT:
+            elif dtype == np.int64 or dtype == np.int32:
                 op.output_type.extend([mace_pb2.DT_INT32])
             else:
                 mace_check(False, "data type %s not supported" % dtype)
@@ -959,7 +955,14 @@ class OnnxConverter(base_converter.ConverterInterface):
                 if len(const_tensor.dims) == 0:
                     value_arg = op.arg.add()
                     value_arg.name = MaceKeyword.mace_scalar_input_str
-                    value_arg.f = const_tensor.float_data[0]
+                    if const_tensor.data_type == mace_pb2.DT_INT32:
+                        value_arg.f = float(const_tensor.int32_data[0])
+                    elif const_tensor.data_type == mace_pb2.DT_FLOAT:
+                        value_arg.f = const_tensor.float_data[0]
+                    else:
+                        mace_check(False,
+                                   "Does not support param's data type %s"
+                                   % const_tensor.data_type)
                     value_index_arg = op.arg.add()
                     value_index_arg.name = \
                         MaceKeyword.mace_scalar_input_index_str
@@ -972,7 +975,14 @@ class OnnxConverter(base_converter.ConverterInterface):
                 if len(const_tensor.dims) == 0:
                     value_arg = op.arg.add()
                     value_arg.name = MaceKeyword.mace_scalar_input_str
-                    value_arg.f = const_tensor.float_data[0]
+                    if const_tensor.data_type == mace_pb2.DT_INT32:
+                        value_arg.f = float(const_tensor.int32_data[0])
+                    elif const_tensor.data_type == mace_pb2.DT_FLOAT:
+                        value_arg.f = const_tensor.float_data[0]
+                    else:
+                        mace_check(False,
+                                   "Does not support param's data type %s"
+                                   % const_tensor.data_type)
                     value_index_arg = op.arg.add()
                     value_index_arg.name = \
                         MaceKeyword.mace_scalar_input_index_str
diff --git a/tools/python/transform/shape_inference.py b/tools/python/transform/shape_inference.py
index e035944d3231094f3dac5707dc775325da02e6be..ac97859bcdc82942c09d053e89f7ac2f2aa9663d 100644
--- a/tools/python/transform/shape_inference.py
+++ b/tools/python/transform/shape_inference.py
@@ -253,7 +253,7 @@ class ShapeInference(object):
         aspect_ratio = ConverterUtil.get_arg(op, MaceKeyword.mace_aspect_ratio_str).floats  # noqa
         num_prior = len(aspect_ratio) * len(min_size) + len(max_size)
 
-        output_shape[2] = num_prior * input_h * input_w * 4
+        output_shape[2] = int(num_prior * input_h * input_w * 4)
         self.add_output_shape(op, [output_shape])
 
     def infer_shape_reshape(self, op):
@@ -275,7 +275,7 @@ class ShapeInference(object):
                     output_shape[i] = dim[i]
                     product *= dim[i]
             if idx != -1:
-                output_shape[idx] = input_size / product
+                output_shape[idx] = int(input_size / product)
             self.add_output_shape(op, [output_shape])
         else:
             output_shape = []
diff --git a/tools/python/transform/transformer.py b/tools/python/transform/transformer.py
index 4c56b233f0f9395588d471168cdf4b350cc603c7..a080c12e904a1e21fd4340313773b976b785a53d 100644
--- a/tools/python/transform/transformer.py
+++ b/tools/python/transform/transformer.py
@@ -1440,15 +1440,15 @@ class Transformer(base_converter.ConverterInterface):
                                 arg.i = 1
                             elif arg.i == 3:
                                 arg.i = 2
-
-                        producer = self._producer[op.input[0]]
-                        input_shape = producer.output_shape[0].dims
-                        if producer.type == MaceOp.FullyConnected.name and \
-                                len(input_shape) == 2:
-                            axis_arg = ConverterUtil.get_arg(
-                                op, MaceKeyword.mace_axis_str)
-                            if axis_arg.i == 1:
-                                axis_arg.i = 3
+                        if op.input[0] in self._producer:
+                            producer = self._producer[op.input[0]]
+                            input_shape = producer.output_shape[0].dims
+                            if (producer.type == MaceOp.FullyConnected.name
+                                    and len(input_shape) == 2):
+                                axis_arg = ConverterUtil.get_arg(
+                                        op, MaceKeyword.mace_axis_str)
+                                if axis_arg.i == 1:
+                                    axis_arg.i = 3
 
             elif op.type == MaceOp.Squeeze.name:
                 for arg in op.arg: