diff --git a/official/cv/densenet/modelarts/train_start.py b/official/cv/densenet/modelarts/train_start.py
index f3a8b7d7071061fd1781fd1cac20af1a7d866dfa..3248f05a64a76aa813e42525d53e85fc08b641a7 100644
--- a/official/cv/densenet/modelarts/train_start.py
+++ b/official/cv/densenet/modelarts/train_start.py
@@ -149,7 +149,7 @@ def _export_air(ckpt_dir):
     param_dict = load_checkpoint(ckpt_file)
     load_param_into_net(net, param_dict)
 
-    input_arr = Tensor(np.zeros([1, 3, 224, 224],
+    input_arr = Tensor(np.zeros([1, 3, config.image_size[0], config.image_size[1]],
                                 np.float32))
     print("Start export air.")
     export(net, input_arr, file_name=config.file_name,
diff --git a/research/cv/RCAN/script/run_ascend_distribute.sh b/research/cv/RCAN/script/run_ascend_distribute.sh
index 2bebe962b0fb34381e79dc7b24c2f426c7890a7c..9f09e5942ebfc25c1c25a3eee762c1357d488737 100644
--- a/research/cv/RCAN/script/run_ascend_distribute.sh
+++ b/research/cv/RCAN/script/run_ascend_distribute.sh
@@ -59,7 +59,7 @@ for ((i = 0; i < ${DEVICE_NUM}; i++)); do
   nohup python train.py \
         --batch_size 16 \
         --lr 1e-4 \
-        --scale 2+3+4 \
+        --scale 2 \
         --task_id 0 \
         --dir_data $PATH2 \
         --epochs 500 \
diff --git a/research/cv/RCAN/src/rcan_model.py b/research/cv/RCAN/src/rcan_model.py
index 6ffbf6680aa9513e1485e2715fbf3f1a77332b02..49517596ca36c1b3e0250b00fef9d5ffad56b09e 100644
--- a/research/cv/RCAN/src/rcan_model.py
+++ b/research/cv/RCAN/src/rcan_model.py
@@ -86,12 +86,11 @@ class Upsampler(nn.Cell):
         """rcan"""
         super(Upsampler, self).__init__()
         m = []
-        for s in scale:
-            if (s & (s - 1)) == 0:
-                for _ in range(int(math.log(s, 2))):
-                    m.append(SmallUpSampler(conv, 2, n_feats, has_bias=has_bias))
-            elif s == 3:
-                m.append(SmallUpSampler(conv, 3, n_feats, has_bias=has_bias))
+        if (scale & (scale - 1)) == 0:
+            for _ in range(int(math.log(scale, 2))):
+                m.append(SmallUpSampler(conv, 2, n_feats, has_bias=has_bias))
+        elif scale == 3:
+            m.append(SmallUpSampler(conv, 3, n_feats, has_bias=has_bias))
         self.net = nn.SequentialCell(m)
 
     def construct(self, x):
@@ -186,7 +185,7 @@ class RCAN(nn.Cell):
         n_feats = args.n_feats
         kernel_size = 3
         reduction = args.reduction
-        scale = args.scale
+        scale = args.scale[0]
         self.dytpe = mstype.float16
 
         # RGB mean for DIV2K