add random forest prior to Bayesian Optimization (#81)

* add random forest prior to Bayesian Optimization

* add citation for BO in readme

* add requirements

README.md

@@ -29,6 +29,7 @@ A Suggestion is an algorithm to construct a parameter set. Currently Katib suppo
 * random
 * grid
 * [hyperband](https://arxiv.org/pdf/1603.06560.pdf)
+* [bayesian optimization](https://arxiv.org/pdf/1012.2599.pdf)
 
 ## Components in Katib
 
@@ -42,6 +43,7 @@ Each component communicates with others via GRPC and the API is defined at `api/
     - vizier-suggestion-random
     - vizier-suggestion-grid
     - vizier-suggestion-hyperband
+    - vizier-suggestion-bayesianoptimization
 - modeldb : WebUI
     - modeldb-frontend
     - modeldb-backend

cmd/suggestion/bayesianoptimization/requirements.txt

@@ -2,3 +2,5 @@ grpcio
 numpy>=1.13.3
 scikit-learn>=0.19.0
 scipy>=0.19.1
+forestci
+protobuf

pkg/suggestion/bayesian_service.py

@@ -69,7 +69,10 @@ def GenerateTrials(self, request, context):
             length_scale=self.service_params[request.study_id]["length_scale"],
             noise=self.service_params[request.study_id]["noise"],
             nu=self.service_params[request.study_id]["nu"],
-            kernel_type=self.service_params[request.study_id]["kernel_type"]
+            kernel_type=self.service_params[request.study_id]["kernel_type"],
+            n_estimators=self.service_params[request.study_id]["n_estimators"],
+            max_features=self.service_params[request.study_id]["max_features"],
+            model_type=self.service_params[request.study_id]["model_type"],
         )
         x_next = alg.get_suggestion().squeeze()
 
@@ -113,7 +116,10 @@ def SetSuggestionParameters(self, request, context):
                 "nu": None,
                 "kernel_type": None,
                 "mode": None,
-                "trade_off": None
+                "trade_off": None,
+                "n_estimators": None,
+                "max_features": None,
+                "model_type": None,
             }
         for param in request.suggestion_parameters:
             if param.name not in self.service_params[request.study_id].keys():
@@ -122,7 +128,7 @@ def SetSuggestionParameters(self, request, context):
                 return api_pb2.SetSuggestionParametersReply()
             if param.name == "length_scale" or param.name == "noise" or param.name == "nu" or param.name == "trade_off":
                 self.service_params[request.study_id][param.name] = float(param.value)
-            elif param.name == "N":
+            elif param.name == "N" or param.name == "n_estimators":
                 self.service_params[request.study_id][param.name] = int(param.value)
             elif param.name == "kernel_type":
                 if param.value != "rbf" and param.value != "matern":
@@ -134,6 +140,10 @@ def SetSuggestionParameters(self, request, context):
                     context.set_code(grpc.StatusCode.UNKNOWN)
                     context.set_details("unknown acquisition mode: " + param.name)
                 self.service_params[request.study_id][param.name] = param.value
+            elif param.name == "model_type":
+                if param.value != "rf" and param.value != "gp":
+                    context.set_code(grpc.StatusCode.UNKNOWN)
+                    context.set_details("unknown model_type: " + param.name)
 
         return api_pb2.SetSuggestionParametersReply()
 

pkg/suggestion/bayesianoptimization/src/acquisition_func/acquisition_func.py

@@ -1,35 +1,53 @@
 """ module for acquisition function"""
 import numpy as np
 from scipy.stats import norm
+import forestci as fci
 
-from pkg.suggestion.bayesianoptimization.src.model.Model import Model
+from pkg.suggestion.bayesianoptimization.src.model.gp import GaussianProcessModel
+from pkg.suggestion.bayesianoptimization.src.model.rf import RandomForestModel
 
 
 class AcquisitionFunc:
     """ class for acquisition function
     expected improvement in this case
     """
-    def __init__(self, X_train, y_train, current_optimal, mode, trade_off, length_scale, noise, nu, kernel_type):
+    def __init__(self, X_train, y_train, current_optimal, mode, trade_off, length_scale,
+                 noise, nu, kernel_type, n_estimators, max_features, model_type):
         """
         :param mode: pi: probability of improvement, ei: expected improvement, lcb: lower confident bound
         :param trade_off: a parameter to control the trade off between exploiting and exploring
+        :param model_type: gp: gaussian process, rf: random forest
         """
         self.X_train = X_train
         self.y_train = y_train
         self.current_optimal = current_optimal
         self.mode = mode or "ei"
         self.trade_off = trade_off or 0.01
-        self.model = Model(
-            length_scale=length_scale,
-            noise=noise,
-            nu=nu,
-            kernel_type=kernel_type,
-        )
+        self.model_type = model_type or "gp"
+        if self.model_type == "gp":
+            self.model = GaussianProcessModel(
+                length_scale=length_scale,
+                noise=noise,
+                nu=nu,
+                kernel_type=kernel_type,
+            )
+        else:
+            self.model = RandomForestModel(
+                n_estimators=n_estimators,
+                max_features=max_features,
+            )
 
     def compute(self, X_test):
-        self.model.gp.fit(self.X_train, self.y_train)
-        y_mean, y_std = self.model.gp.predict(X_test, return_std=True)
-        y_variance = y_std ** 2
+        if self.model_type == "gp":
+            self.model.gp.fit(self.X_train, self.y_train)
+            y_mean, y_std = self.model.gp.predict(X_test, return_std=True)
+            y_variance = y_std ** 2
+        else:
+            self.model.rf.fit(self.y_train, self.y_train)
+            y_mean = self.model.rf.predict(X_test)
+            y_variance = fci.random_forest_error(self.model.rf, self.X_train, X_test)
+            y_std = np.sqrt(y_variance)
+
         z = (y_mean - self.current_optimal - self.trade_off) / y_std
 
         if self.mode == "ei":

pkg/suggestion/bayesianoptimization/src/bayesian_optimization_algorithm.py

@@ -8,7 +8,7 @@
 class BOAlgorithm:
     """ class for bayesian optimization """
     def __init__(self, dim, N, lowerbound, upperbound, X_train, y_train, mode, trade_off,
-                 length_scale, noise, nu, kernel_type):
+                 length_scale, noise, nu, kernel_type, n_estimators, max_features, model_type):
         # np.random.seed(0)
         self.dim = dim
         self.N = N or 100
@@ -44,6 +44,9 @@ def __init__(self, dim, N, lowerbound, upperbound, X_train, y_train, mode, trade
             noise=noise,
             nu=nu,
             kernel_type=kernel_type,
+            n_estimators=n_estimators,
+            max_features=max_features,
+            model_type=model_type,
         )
 
     def get_suggestion(self):

pkg/suggestion/bayesianoptimization/src/global_optimizer/global_optimizer.py

@@ -70,7 +70,7 @@ class GlobalOptimizer:
     """ class for the global optimizer """
 
     def __init__(self, N, l, u, scaler, X_train, y_train, current_optimal, mode, trade_off, length_scale,
-                 noise, nu, kernel_type):
+                 noise, nu, kernel_type, n_estimators, max_features, model_type):
         self.N = N
         self.l = l
         self.u = u
@@ -87,6 +87,9 @@ def __init__(self, N, l, u, scaler, X_train, y_train, current_optimal, mode, tra
             noise=noise,
             nu=nu,
             kernel_type=kernel_type,
+            n_estimators=n_estimators,
+            max_features=max_features,
+            model_type=model_type,
         )
 
     def potential_opt(self, f_min):

pkg/suggestion/bayesianoptimization/src/model/Model.py → pkg/suggestion/bayesianoptimization/src/model/gp.py

@@ -3,7 +3,7 @@
 from sklearn.gaussian_process import GaussianProcessRegressor
 
 
-class Model:
+class GaussianProcessModel:
     """ use the gaussian process as a prior """
     def __init__(self, length_scale, noise, nu, kernel_type):
         """

pkg/suggestion/bayesianoptimization/src/model/rf.py

@@ -0,0 +1,11 @@
+from sklearn.ensemble import RandomForestRegressor
+
+
+class RandomForestModel:
+    def __init__(self, n_estimators, max_features):
+        n_estimators = n_estimators or 50
+        max_features = max_features or "auto"
+        self.rf = RandomForestRegressor(
+            n_estimators=n_estimators,
+            max_features=max_features,
+        )

pkg/suggestion/test_client.py

@@ -27,7 +27,15 @@ def run():
             api_pb2.SuggestionParameter(
                 name="trade_off",
                 value="0.01",
-            )
+            ),
+            api_pb2.SuggestionParameter(
+                name="model_type",
+                value="gp",
+            ),
+            api_pb2.SuggestionParameter(
+                name="n_estimators",
+                value="50",
+            ),
         ]
     ))
     completed_trials = []