fix quniform issue ; change related doc

docs/en_US/Tutorial/SearchSpaceSpec.md

@@ -45,7 +45,8 @@ All types of sampling strategies and their parameter are listed here:
   * When optimizing, this variable is constrained to a two-sided interval.
 
 * {"_type":"quniform","_value":[low, high, q]}
-  * Which means the variable value is a value like round(uniform(low, high) / q) * q
+  * Which means the variable value is a value like randint(np.floor((high-low)/q)+1)*q + low. For example, for _value specified as [0, 10, 2.5], possible values are [0, 2.5, 5.0, 7.5, 10.0]; For _value specified as [2, 10, 5], possible values are [2, 7]. All possible values have the same possibility to be selected.   
+
   * Suitable for a discrete value with respect to which the objective is still somewhat "smooth", but which should be bounded both above and below. If you want to uniformly choose integer from a range [low, high], you can write `_value` like this: `[low, high, 1]`.
 
 * {"_type":"loguniform","_value":[low, high]}

src/sdk/pynni/nni/parameter_expressions.py

@@ -57,7 +57,7 @@ def quniform(low, high, q, random_state):
     q: sample step
     random_state: an object of numpy.random.RandomState
     '''
-    return np.round(uniform(low, high, random_state) / q) * q
+    return randint(np.floor((high-low)/q)+1, random_state) * q + low
 
 
 def loguniform(low, high, random_state):