Support convert nested dict in TPE (#1023)

docs/en_US/Builtin_Tuner.md

@@ -132,6 +132,8 @@ tuner:
 
 > Builtin Tuner Name: **SMAC**
 
+**Please note that SMAC doesn't support running on windows currently. The specific reason can be referred to this [github issue](https://github.com/automl/SMAC3/issues/483).**
+
 **Installation**
 
 SMAC need to be installed by following command before first use.

examples/trials/mnist-cascading-search-space/mnist.py

@@ -145,7 +145,7 @@ def parse_init_json(data):
         if value == 'Empty':
             params[key] = ['Empty']
         else:
-            params[key] = [value[0], value[1]['_value'], value[1]['_value']]
+            params[key] = [value[0], value[1], value[1]]
     return params
 
 

examples/trials/mnist-cascading-search-space/sample.json

-{"layer2": "Empty", "layer8": ["Conv", {"_index": 0, "_value": 2}], "layer3": ["Avg_pool", {"_index": 2,     "_value": 5}], "layer0": ["Max_pool", {"_index": 2, "_value": 5}], "layer1": ["Conv", {"_index": 0, "_value": 2}], "layer6": ["Max_pool",     {"_index": 1, "_value": 3}], "layer7": ["Max_pool", {"_index": 2, "_value": 5}], "layer9": ["Conv", {"_index": 0, "_value": 2}], "layer4":     ["Avg_pool", {"_index": 1, "_value": 3}], "layer5": ["Avg_pool", {"_index": 2, "_value": 5}]}
+{
+    "layer2": "Empty", 
+    "layer8": ["Conv",  2], 
+    "layer3": ["Avg_pool", 5], 
+    "layer0": ["Max_pool", 5], 
+    "layer1": ["Conv", 2], 
+    "layer6": ["Max_pool", 3], 
+    "layer7": ["Max_pool", 5], 
+    "layer9": ["Conv", 2], 
+    "layer4": ["Avg_pool", 3], 
+    "layer5": ["Avg_pool", 5]
+}

src/sdk/pynni/nni/hyperopt_tuner/hyperopt_tuner.py

@@ -139,40 +139,50 @@ def json2vals(in_x, vals, out_y, name=ROOT):
         for i, temp in enumerate(in_x):
             json2vals(temp, vals[i], out_y, name + '[%d]' % i)
 
-def params2tuner_params(in_x, parameter):
+def _add_index(in_x, parameter):
     """
-    change parameters in NNI format to parameters in hyperopt format.
-    For example, NNI receive parameters like:
+    change parameters in NNI format to parameters in hyperopt format(This function also support nested dict.).
+    For example, receive parameters like:
         {'dropout_rate': 0.8, 'conv_size': 3, 'hidden_size': 512}
     Will change to format in hyperopt, like:
         {'dropout_rate': 0.8, 'conv_size': {'_index': 1, '_value': 3}, 'hidden_size': {'_index': 1, '_value': 512}}
     """
-    tuner_params = dict()
-    for key in parameter.keys():
-        value = parameter[key]
-        _type = in_x[key][TYPE]
-        if _type == 'choice':
-            _idx = in_x[key][VALUE].index(value)
-            tuner_params[key] = {
-                INDEX: _idx,
-                VALUE: value
-            }
+    if TYPE not in in_x: # if at the top level
+        out_y = dict()
+        for key, value in parameter.items():
+            out_y[key] = _add_index(in_x[key], value)
+        return out_y
+    elif isinstance(in_x, dict):
+        value_type = in_x[TYPE]
+        value_format = in_x[VALUE]
+        if value_type == "choice":
+            choice_name = parameter[0] if isinstance(parameter, list) else parameter
+            for pos, item in enumerate(value_format): # here value_format is a list
+                if isinstance(item, list): # this format is ["choice_key", format_dict]
+                    choice_key = item[0]
+                    choice_value_format = item[1]
+                    if choice_key == choice_name:
+                        return {INDEX: pos, VALUE: [choice_name, _add_index(choice_value_format, parameter[1])]}
+                elif choice_name == item:
+                    return {INDEX: pos, VALUE: item}
         else:
-            tuner_params[key] = value
-    return tuner_params
+            return parameter
 
 def _split_index(params):
     """
     Delete index infromation from params
     """
-    result = {}
-    for key in params:
-        if isinstance(params[key], dict):
-            value = params[key][VALUE]
-        else:
-            value = params[key]
-        result[key] = value
-    return result
+    if isinstance(params, list):
+        return [params[0], _split_index(params[1])]
+    elif isinstance(params, dict):
+        if INDEX in params.keys():
+            return _split_index(params[VALUE])
+        result = dict()
+        for key in params:
+            result[key] = _split_index(params[key])
+        return result
+    else:
+        return params
 
 
 class HyperoptTuner(Tuner):
@@ -399,6 +409,6 @@ class HyperoptTuner(Tuner):
                 continue
             self.supplement_data_num += 1
             _parameter_id = '_'.join(["ImportData", str(self.supplement_data_num)])
-            self.total_data[_parameter_id] = params2tuner_params(self.json, _params)
+            self.total_data[_parameter_id] = _add_index(in_x=self.json, parameter=_params)
             self.receive_trial_result(parameter_id=_parameter_id, parameters=_params, value=_value)
         logger.info("Successfully import data to TPE/Anneal tuner.")