diff --git a/tools/image/image_to_tensor.py b/tools/image/image_to_tensor.py
index 936d8f3972cc1be4112fbda1faab03550e03580f..d39c07c3e32f6b2f635549d621355ea1c002080c 100644
--- a/tools/image/image_to_tensor.py
+++ b/tools/image/image_to_tensor.py
@@ -3,6 +3,8 @@ import os
 import sys
 import tensorflow as tf
 
+# TODO(liyin): use dataset api and estimator with distributed strategy
+
 FLAGS = None
 
 
@@ -32,32 +34,38 @@ def parse_args():
 
 
 def images_to_tensors(input_files, image_shape, mean_values=None):
-    for i in xrange(len(input_files)):
-        with tf.Session() as sess:
-            with tf.gfile.FastGFile(input_files[i], 'rb') as f:
-                image_data = f.read()
-                image_data = tf.image.decode_image(image_data,
-                                                   channels=image_shape[2])
-                if mean_values:
-                    image_data = tf.cast(image_data, dtype=tf.float32)
-                    mean_tensor = tf.constant(mean_values, dtype=tf.float32,
-                                              shape=[1, 1, image_shape[2]])
-                    image_data = (image_data - mean_tensor) / 255.0
-                else:
-                    image_data = tf.image.convert_image_dtype(image_data,
-                                                              dtype=tf.float32)
-                    image_data = tf.subtract(image_data, 0.5)
-                    image_data = tf.multiply(image_data, 2.0)
+    with tf.Graph().as_default():
+        image_data = tf.placeholder(tf.string, name='input')
+        image_data = tf.image.decode_image(image_data,
+                                           channels=image_shape[2])
+        if mean_values:
+            image_data = tf.cast(image_data, dtype=tf.float32)
+            mean_tensor = tf.constant(mean_values, dtype=tf.float32,
+                                      shape=[1, 1, image_shape[2]])
+            image_data = (image_data - mean_tensor) / 255.0
+        else:
+            image_data = tf.image.convert_image_dtype(image_data,
+                                                      dtype=tf.float32)
+            image_data = tf.subtract(image_data, 0.5)
+            image_data = tf.multiply(image_data, 2.0)
 
-                image_data = tf.expand_dims(image_data, 0)
-                image_data = tf.image.resize_bilinear(image_data,
-                                                      image_shape[:2],
-                                                      align_corners=False)
+        image_data = tf.expand_dims(image_data, 0)
+        image_data = tf.image.resize_bilinear(image_data,
+                                              image_shape[:2],
+                                              align_corners=False)
 
-                image = sess.run(image_data)
-                output_file = os.path.join(FLAGS.output_dir, os.path.splitext(
-                    os.path.basename(input_files[i]))[0] + '.dat')
-                image.tofile(output_file)
+        with tf.Session() as sess:
+            for i in xrange(len(input_files)):
+                with tf.gfile.FastGFile(input_files[i], 'rb') as f:
+                    src_image = f.read()
+                    dst_image = sess.run(image_data,
+                                         feed_dict={'input:0': src_image})
+                    output_file = os.path.join(FLAGS.output_dir,
+                                               os.path.splitext(
+                                                   os.path.basename(
+                                                       input_files[i]))[
+                                                   0] + '.dat')
+                    dst_image.tofile(output_file)
 
 
 def main(unused_args):
diff --git a/tools/image/tensor_to_image.py b/tools/image/tensor_to_image.py
index dc903f0132e1b29aca3544fc0082e04ad7e14428..ce18628eae52c98ada0c4abb06ffc61be836d640 100644
--- a/tools/image/tensor_to_image.py
+++ b/tools/image/tensor_to_image.py
@@ -4,6 +4,8 @@ import sys
 import numpy as np
 import tensorflow as tf
 
+# TODO(liyin): use dataset api and estimator with distributed strategy
+
 FLAGS = None
 
 
@@ -27,22 +29,26 @@ def parse_args():
 
 
 def tensors_to_images(input_files, image_shape):
-    for i in xrange(len(input_files)):
+    with tf.Graph().as_default():
+        input = tf.placeholder(tf.float32, shape=image_shape, name='input')
+        output = tf.placeholder(tf.string, name='output_file')
+        # use the second channel if it is gray image
+        if image_shape[2] == 2:
+            _, input = tf.split(input, 2, axis=2)
+        tensor_data = tf.image.convert_image_dtype(input,
+                                                   tf.uint8,
+                                                   saturate=True)
+        image_data = tf.image.encode_jpeg(tensor_data, quality=100)
+        writer = tf.write_file(output, image_data, name='output_writer')
+
         with tf.Session() as sess:
-            tensor_data = np.fromfile(input_files[i], dtype=np.float32) \
-                .reshape(image_shape)
-            # use the second channel if it is gray image
-            if image_shape[2] == 2:
-                _, tensor_data = tf.split(tensor_data, 2, axis=2)
-            tensor_data = tf.image.convert_image_dtype(tensor_data,
-                                                       tf.uint8,
-                                                       saturate=True)
-            image_data = tf.image.encode_jpeg(tensor_data, quality=100)
-            image = sess.run(image_data)
-            output_file = os.path.join(FLAGS.output_dir, os.path.splitext(
-                os.path.basename(input_files[i]))[0] + '.jpg')
-            writer = tf.write_file(output_file, image)
-            sess.run(writer)
+            for i in xrange(len(input_files)):
+                input_data = np.fromfile(input_files[i], dtype=np.float32) \
+                    .reshape(image_shape)
+                output_file = os.path.join(FLAGS.output_dir, os.path.splitext(
+                    os.path.basename(input_files[i]))[0] + '.jpg')
+                sess.run(writer, feed_dict={'input:0': input_data,
+                                            'output_file:0': output_file})
 
 
 def main(unused_args):