diff --git a/reinforcement_learning/common/sagemaker_rl/coach_launcher.py b/reinforcement_learning/common/sagemaker_rl/coach_launcher.py
index 065c745a17..79a6c7a6fa 100644
--- a/reinforcement_learning/common/sagemaker_rl/coach_launcher.py
+++ b/reinforcement_learning/common/sagemaker_rl/coach_launcher.py
@@ -8,7 +8,6 @@
 import rl_coach.core_types 
 from rl_coach import logger
 from rl_coach.logger import screen
-import tensorflow as tf
 import argparse
 import copy
 import logging
@@ -192,6 +191,7 @@ def _save_tf_model(self):
         ckpt_dir = '/opt/ml/output/data/checkpoint'
         model_dir = '/opt/ml/model'
 
+        import tensorflow as tf # importing tensorflow here so that MXNet docker image is compatible with this file.
         # Re-Initialize from the checkpoint so that you will have the latest models up.
         tf.train.init_from_checkpoint(ckpt_dir,
                                       {'main_level/agent/online/network_0/': 'main_level/agent/online/network_0'})