Lazy initialize mnist dataset.

Fix unittest

python/paddle/v2/dataset/mnist.py

@@ -16,18 +16,29 @@ def __mnist_reader_creator__(data, target):
 
 
 TEST_SIZE = 10000
+X_train = None
+X_test = None
+y_train = None
+y_test = None
 
-data = sklearn.datasets.mldata.fetch_mldata(
-    "MNIST original", data_home=DATA_HOME)
-X_train, X_test, y_train, y_test = sklearn.model_selection.train_test_split(
-    data.data, data.target, test_size=TEST_SIZE, random_state=0)
+
+def __initialize_dataset__():
+    global X_train, X_test, y_train, y_test
+    if X_train is not None:
+        return
+    data = sklearn.datasets.mldata.fetch_mldata(
+        "MNIST original", data_home=DATA_HOME)
+    X_train, X_test, y_train, y_test = sklearn.model_selection.train_test_split(
+        data.data, data.target, test_size=TEST_SIZE, random_state=0)
 
 
 def train_creator():
+    __initialize_dataset__()
     return __mnist_reader_creator__(X_train, y_train)
 
 
 def test_creator():
+    __initialize_dataset__()
     return __mnist_reader_creator__(X_test, y_test)