diff --git a/src/communication.py b/src/communication.py
index 1b9bd42..99fa584 100644
--- a/src/communication.py
+++ b/src/communication.py
@@ -223,11 +223,11 @@ class ClusterCommunicationModule():
                     docker.types.DeviceRequest(count=-1, capabilities=[['gpu']])
                 ],
                 name=f'train-{train_args["index"]}',
-                env={
+                environment={
                     'GPU_NUM': self.node_manager.GPU_num,
                     'NODE_NUM': train_args['node_num'],
                     'NODE_RANK': train_args['index'],
-                    'MASTER_IP': 'train-0',
+                    'MASTER_IP': 'train-0' if self.node_manager.status == 'worker' else '127.0.0.1',
                     'MASTER_PORT': 21046,
                 },
                 detach=True
@@ -241,7 +241,7 @@ class ClusterCommunicationModule():
         status_code = result['StatusCode']
         print(status_code, type(status_code))
 
-    def scatter_container(self, image_name, train=False):
+    def scatter_container(self, image_name, train):
         def master_run(image_name):
             print("[Master] run")
             train_args = {
diff --git a/src/node_manager.py b/src/node_manager.py
index 1d0d5c9..460b158 100644
--- a/src/node_manager.py
+++ b/src/node_manager.py
@@ -120,9 +120,13 @@ class NodeManager():
             '''
 
             data_image = "snsd0805/cifar100-dataset:v1"
+            train_image = "snsd0805/cifar100-train:v3"
 
             # Start Downloading
-            self.cluster_communication_module.scatter_container(data_image)
+            # self.cluster_communication_module.scatter_container(data_image, train=False)
+
+            # start training
+            self.cluster_communication_module.scatter_container(train_image, train=True)
 
 
         else: