fix reamde

distill_knowledge/README.cn.md

@@ -109,7 +109,7 @@ def mlp_student(img, drop_prob, h_size):
 | StudentNet 30units with 4.0temp soft targets | 97.01%  |
 
 训练过程中的测试集准确率如下图所示，可以看到使用soft targets训练的小网络的收敛速度快于不使用soft targets的网络。
-![收敛](https://github.com/likesiwell/models/blob/develop/distill_knowledge/images/plots.png)
+![收敛](https://github.com/likesiwell/models/blob/distill-branch/distill_knowledge/images/plots.png)
 
 
 ## 参考文献

distill_knowledge/README.md

@@ -110,7 +110,7 @@ When we train the small network with soft targets of temperature 4.0, we get the
 | StudentNet 30units with 4.0temp soft targets | 97.01%  |
 
 The testing accuracy during training procedure are visualized here. We can see that using soft targets speed up the convergence.
-![convergence](https://github.com/likesiwell/models/blob/develop/distill_knowledge/images/plots.png)
+![convergence](https://github.com/likesiwell/models/blob/distill-branch/distill_knowledge/images/plots.png)
 
 
 

distill_knowledge/mnist_prepare.py

 import paddle
 import numpy as np
+import os
 reader = paddle.dataset.mnist.train()
 img_list = []
 label_list = []
@@ -23,6 +24,11 @@ test_x = np.vstack(img_list)
 test_y = np.vstack(label_list)
 print(test_x.shape, test_y.shape)
 
+if not os.path.exits('./data/'):
+    os.makedirs('./data/')
+if not os.path.exits('./models/'):
+    os.makedirs('./models')
+
 np.savez(
     './data/mnist.npz',
     train_x=train_x,