[WIP] Implement Tree proposal with Dice

discretesampling/base/random.py

@@ -1,5 +1,6 @@
 from random import randint
 import random
+import numpy as np
 
 
 class Random(object):
@@ -35,6 +36,31 @@ def eval(self):
         return random.choices(population=self.population, weights=self.weights, cum_weights=self.cum_weights, k=self.k)
 
 
+class Dice(object):
+    def __init__(self, probabilities, outcomes):
+        assert len(outcomes) == len(probabilities), "Invalid PMF specified, x and p" +\
+             " of different lengths"
+        probabilities = np.array(probabilities)
+        tolerance = np.sqrt(np.finfo(np.float64).eps)
+        assert abs(1 - sum(probabilities)) < tolerance, "Invalid PMF specified," +\
+            " sum of probabilities !~= 1.0"
+        assert all(probabilities >= 0.), "Invalid PMF specified, all probabilities" +\
+            " must be > 0"
+        self.probabilities = probabilities
+        self.outcomes = outcomes
+        self.pmf = probabilities
+        self.cmf = np.cumsum(probabilities)
+        self.randomiser = Random()
+
+    def eval(self):
+        q = self.randomiser.eval()
+        x = self.outcomes[np.argmax(self.cmf >= q)]
+
+        while callable(x):
+            x = x()
+        return x
+
+
 def set_seed(seed):
     """
     :param seed: random seed

discretesampling/base/types.py

@@ -28,19 +28,8 @@ def getOptimalLKernelType(self):
 
 
 class DiscreteVariableProposal:
-    def __init__(self, values, probs):
-        # Check dims and probs are valid
-        assert len(values) == len(probs), "Invalid PMF specified, x and p" +\
-             " of different lengths"
-        probs = np.array(probs)
-        tolerance = np.sqrt(np.finfo(np.float64).eps)
-        assert abs(1 - sum(probs)) < tolerance, "Invalid PMF specified," +\
-            " sum of probabilities !~= 1.0"
-        assert all(probs > 0), "Invalid PMF specified, all probabilities" +\
-            " must be > 0"
-        self.x = values
-        self.pmf = probs
-        self.cmf = np.cumsum(probs)
+    def __init__(self, moves_dice):
+        self.moves_dice = moves_dice
 
     @classmethod
     def norm(self, x):
@@ -54,8 +43,7 @@ def heuristic(self, x, y):
         return True
 
     def sample(self):
-        q = random.random()  # random unif(0,1)
-        return self.x[np.argmax(self.cmf >= q)]
+        return self.moves_dice.eval()
 
     def eval(self, y):
         try:

discretesampling/domain/decision_tree/tree_distribution.py

@@ -1,11 +1,14 @@
 import numpy as np
 from math import log, inf
 import copy
-from ...base.random import Random
+from ...base.random import Dice
 from ...base import types
 
 
 class TreeProposal(types.DiscreteVariableProposal):
+    moves_prob = [0.4, 0.1, 0.1, 0.4]
+    moves = ["prune", "swap", "change", "grow"]  # noqa
+
     def __init__(self, tree):
         self.X_train = tree.X_train
         self.y_train = tree.y_train
@@ -21,40 +24,25 @@ def norm(self, tree):
     def heuristic(self, x, y):
         return y < x or abs(x-y) < 2
 
-    def sample(self):
-        # initialise the probabilities of each move
-        moves = ["prune", "swap", "change", "grow"]  # noqa
-        moves_prob = [0.4, 0.1, 0.1, 0.4]
+    def get_moves_prob(self):
         if len(self.tree.tree) == 1:
             moves_prob = [0.0, 0.0, 0.5, 0.5]
-        moves_probabilities = np.cumsum(moves_prob)
-        random_number = Random().eval()
-        newTree = copy.deepcopy(self.tree)
-        if random_number < moves_probabilities[0]:
-            # prune
-            newTree = newTree.prune()
-
-        elif random_number < moves_probabilities[1]:
-            # swap
-            newTree = newTree.swap()
-
-        elif random_number < moves_probabilities[2]:
-            # change
-            newTree = newTree.change()
-
         else:
-            # grow
-            newTree = newTree.grow()
+            moves_prob = self.moves_prob
+        return moves_prob
 
-        return newTree
+    def sample(self):
+        # initialise the probabilities of each move
+        moves_prob = self.get_moves_prob()
+        newTree = copy.deepcopy(self.tree)
+        moves_dice = Dice(moves_prob, [newTree.prune, newTree.swap, newTree.change, newTree.grow])
+
+        return moves_dice.eval()
 
     def eval(self, sampledTree):
         initialTree = self.tree
-        moves_prob = [0.4, 0.1, 0.1, 0.4]
         logprobability = -inf
-        if len(initialTree.tree) == 1:
-            moves_prob = [0.0, 0.0, 0.5, 0.5]
-
+        moves_prob = self.get_moves_prob()
         nodes_differences = [i for i in sampledTree.tree + initialTree.tree
                              if i not in sampledTree.tree or
                              i not in initialTree.tree]