Merge pull request #231 from jklaise/python3-routers

Update epsilon-greedy example to Python 3

examples/routers/epsilon_greedy/EpsilonGreedy.py

@@ -1,7 +1,7 @@
 import random
 import numpy as np
 
-__version__ = "v1.1"
+__version__ = "v1.2"
 
 def n_success_failures(features,reward):
     n_predictions = features.shape[0]
@@ -10,9 +10,9 @@ def n_success_failures(features,reward):
     return n_success, n_failures
 
 class EpsilonGreedy(object):
-    
+
     def __init__(self,n_branches=None,epsilon=0.1,verbose=False):
-        print "Starting Epsilon Greedy Microservice, version {}".format(__version__)
+        print("Starting Epsilon Greedy Microservice, version {}".format(__version__))
         if n_branches is None:
             raise Exception("n_branches parameter must be given")
         self.verbose = verbose
@@ -22,29 +22,29 @@ def __init__(self,n_branches=None,epsilon=0.1,verbose=False):
         self.branches_tries = [0 for _ in range(n_branches)]
         self.n_branches = n_branches
         if self.verbose:
-            print "Router initialised"
-            print "# branches:",self.n_branches
-            print "Epsilon:",self.epsilon
-            print
-        
+            print("Router initialised")
+            print("# branches:",self.n_branches)
+            print("Epsilon:",self.epsilon)
+            print()
+
     def route(self,features,feature_names):
         x = random.random()
         best_branch = self.best_branch
         other_branches = [i for i in range(self.n_branches) if i!=best_branch]
         selected_branch = best_branch if x>self.epsilon else random.choice(other_branches)
         if self.verbose:
-            print "Routing"
-            print "Current best branch:",best_branch
-            print "Selected branch:",selected_branch
-            print
+            print("Routing")
+            print("Current best branch:",best_branch)
+            print("Selected branch:",selected_branch)
+            print()
         return selected_branch
-        
+
     def send_feedback(self,features,feature_names,routing,reward,truth):
         if self.verbose:
-            print "Training"
-            print "Prev success #", self.branches_success
-            print "Prev tries #", self.branches_tries
-            print "Prev best branch:", self.best_branch
+            print("Training")
+            print("Prev success #", self.branches_success)
+            print("Prev tries #", self.branches_tries)
+            print("Prev best branch:", self.best_branch)
         n_success, n_failures = n_success_failures(features,reward)
         self.branches_success[routing] += n_success
         self.branches_tries[routing] += n_success + n_failures
@@ -55,8 +55,7 @@ def send_feedback(self,features,feature_names,routing,reward,truth):
         ]
         self.best_branch = np.argmax(perfs)
         if self.verbose:
-            print "New success #", self.branches_success
-            print "New tries #", self.branches_tries
-            print "New best branch:",self.best_branch
-            print
-
+            print("New success #", self.branches_success)
+            print("New tries #", self.branches_tries)
+            print("New best branch:",self.best_branch)
+            print()

examples/routers/epsilon_greedy/README.md

@@ -1,10 +1,22 @@
 # Epsilon Greedy Router
 
+## Description
+
+An epsilon-greedy router implements a [multi-armed bandit strategy](https://en.wikipedia.org/wiki/Multi-armed_bandit#Semi-uniform_strategies) in which, when presented with *n* models to make predictions, the currently
+best performing model is selected with probability *1-e* while a random model is selected with probability *e*.
+This strategy ensures sending traffic to the best performing model most of the time (exploitation) while allowing for
+some evaluation of other models (exploration). A typical parameter value could be *e=0.1*, but this will depend on the
+desired trade-off between exploration and exploitation.
+
+Note that in this implementation the parameter value *e* is static, but a related strategy called *epsilon-decreasing*
+would see the value of *e* decrease as the number of predictions increases, resulting in a highly explorative behaviour
+at the start and increasingly exploitative behaviour as time goes on.
+
 
 ## Wrap using s2i
 
 ```bash
-s2i build . seldonio/seldon-core-s2i-python2 egreedy-router
+s2i build . seldonio/seldon-core-s2i-python3 egreedy-router
 ```
 
 ## Smoke Test
@@ -20,4 +32,8 @@ Send a data request.
 ```bash
 data='{"data":{"names":["a","b"],"ndarray":[[1.0,2.0]]}}'
 curl -d "json=${data}" http://0.0.0.0:5000/route
-```
+```
+
+## Running on Seldon
+An end-to-end example deploying an epsilon-greedy router to route traffic to 3 models in parallel is available [here](
+https://github.com/SeldonIO/seldon-core/blob/master/notebooks/epsilon_greedy_gcp.ipynb).

examples/routers/epsilon_greedy/requirements.txt

@@ -1 +1 @@
-numpy==1.11.2
+numpy==1.14.3