diff --git a/PaddleCV/image_classification/ResNeXt101_vd_32x4d_local.sh b/PaddleCV/image_classification/ResNeXt101_vd_32x4d_local.sh
new file mode 100644
index 0000000000000000000000000000000000000000..ff2dcdda81f85859a772583f899629c128a2cec9
--- /dev/null
+++ b/PaddleCV/image_classification/ResNeXt101_vd_32x4d_local.sh
@@ -0,0 +1,23 @@
+#Training details
+#Missed
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+export FLAGS_fast_eager_deletion_mode=1
+export FLAGS_eager_delete_tensor_gb=0.0
+export FLAGS_fraction_of_gpu_memory_to_use=0.98
+
+#ResNeXt101_vd_32x4d
+python -m paddle.distributed.launch \
+        --use_paddlecloud \
+        --selected_gpus="0,1,2,3,4,5,6,7" \
+        --log_dir=mylog \
+        train.py \
+	      --model=ResNeXt101_vd_32x4d \
+        --batch_size=256 \
+        --lr_strategy=cosine_decay \
+        --lr=0.1 \
+        --num_epochs=200 \
+        --model_save_dir=output/ \
+        --l2_decay=1e-4 \
+        --use_mixup=True \
+        --use_label_smoothing=True \
+        --label_smoothing_epsilon=0.1
diff --git a/PaddleCV/image_classification/reader.py b/PaddleCV/image_classification/reader.py
index 239f9eae585b23f5b7c0f6b0d0aa912676f3d6e7..4789596fadbe5d1c8257ea5852baeb31d061d46f 100644
--- a/PaddleCV/image_classification/reader.py
+++ b/PaddleCV/image_classification/reader.py
@@ -275,8 +275,7 @@ class ImageNetReader:
             batch_size = 1
         else:
             if settings.use_gpu:
-                batch_size = settings.batch_size // paddle.fluid.core.get_cuda_device_count(
-                )
+                batch_size = settings.batch_size // num_trainers
             else:
                 batch_size = settings.batch_size // int(
                     os.environ.get('CPU_NUM', 1))
@@ -296,7 +295,6 @@ class ImageNetReader:
                             full_lines)
                     elif shuffle:
                         np.random.shuffle(full_lines)
-
                 batch_data = []
                 for line in full_lines:
                     img_path, label = line.split()
@@ -362,10 +360,10 @@ class ImageNetReader:
 
         if settings.use_mixup == True:
             reader = create_mixup_reader(settings, reader)
+            num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
             reader = fluid.io.batch(
                 reader,
-                batch_size=int(settings.batch_size /
-                               paddle.fluid.core.get_cuda_device_count()),
+                batch_size = settings.batch_size // num_trainers,
                 drop_last=True)
         return reader
 
diff --git a/PaddleCV/image_classification/train.py b/PaddleCV/image_classification/train.py
index b11dab999e0f9cf7e61fcd81c869caac7b0e57ad..c45ff12ecdb618ecd9959321e2f04fb9a3c3fa92 100755
--- a/PaddleCV/image_classification/train.py
+++ b/PaddleCV/image_classification/train.py
@@ -29,6 +29,9 @@ from utils import *
 import models
 from build_model import create_model
 
+from paddle.fluid.incubate.fleet.collective import fleet, DistributedStrategy  # new line 1
+from paddle.fluid.incubate.fleet.base import role_maker # new line 2
+
 
 def build_program(is_train, main_prog, startup_prog, args):
     """build program, and add grad op in program accroding to different mode
@@ -62,12 +65,24 @@ def build_program(is_train, main_prog, startup_prog, args):
             # add backward op in program
             if is_train:
                 optimizer = create_optimizer(args)
-                avg_cost = loss_out[0]
-                optimizer.minimize(avg_cost)
                 #XXX: fetch learning rate now, better implement is required here. 
                 global_lr = optimizer._global_learning_rate()
                 global_lr.persistable = True
                 loss_out.append(global_lr)
+                avg_cost = loss_out[0]
+
+                #################################
+                # configure DistributedStrategy #
+                #################################
+                dist_strategy = DistributedStrategy()
+                dist_strategy.nccl_comm_num = 2
+                exec_strategy = fluid.ExecutionStrategy()
+                exec_strategy.num_threads = 3
+                exec_strategy.num_iteration_per_drop_scope = 30
+                dist_strategy.exec_strategy = exec_strategy
+                dist_strategy.fuse_all_reduce_ops = True
+                optimizer = fleet.distributed_optimizer(optimizer, strategy=dist_strategy) # new line 5
+                optimizer.minimize(avg_cost)
                 if args.use_ema:
                     global_steps = fluid.layers.learning_rate_scheduler._decay_step_counter(
                     )
@@ -120,6 +135,9 @@ def train(args):
     Args:
         args: all arguments.    
     """
+    role = role_maker.PaddleCloudRoleMaker(is_collective=True) # new line 3
+    fleet.init(role) # new line 4
+
     startup_prog = fluid.Program()
     train_prog = fluid.Program()
     test_prog = fluid.Program()
@@ -176,8 +194,7 @@ def train(args):
         train_data_loader.set_sample_list_generator(train_reader, places)
         test_data_loader.set_sample_list_generator(test_reader, place)
 
-    compiled_train_prog = best_strategy_compiled(args, train_prog,
-                                                 train_fetch_vars[0], exe)
+    compiled_train_prog = fleet.main_program # change line 1
     #NOTE: this for benchmark
     total_batch_num = 0
     for pass_id in range(args.num_epochs):