Merge pull request #286 from ezCGP/feature/simgan-support

Feature/simgan support

codes/block_definitions/utilities/operators_pytorch.py

@@ -551,8 +551,8 @@ def feature_extraction(input_shapes, *args):
     NOTE: an idea was to make the output of this method to be a 2nd input into the
     discriminator block instead of a primitive of that block
     '''
-    from misc import features
-    class FeatureExtractor(MimicPyTorchModule, features.FeatureExtractor):
+    from codes.utilities import simgan_feature_eval
+    class FeatureExtractor(MimicPyTorchModule, simgan_feature_eval.FeatureExtractor):
         def __init__(self, input_shapes, *args):
             '''
             if input_shapes[0] doesn't 92 features then the code will
@@ -563,7 +563,7 @@ def __init__(self, input_shapes, *args):
             fake_shape = input_shapes[0]
             fake_shape = fake_shape[:-1] + (92,)
             MimicPyTorchModule.__init__(self, [fake_shape], ftn=None)
-            features.FeatureExtractor.__init__(self, *args)
+            simgan_feature_eval.FeatureExtractor.__init__(self, *args)
 
         def get_out_shape(self):
             return (self.input_shapes[0][0], self.num_features)

codes/factory.py

@@ -55,8 +55,8 @@ def build_population(self,
         '''
         my_population = PopulationDefinition()
 
-        if (node_rank==0) and (problem.hall_of_fame_size is not None):
-            my_population.setup_hall_of_fame(problem.hall_of_fame_size)
+        if (node_rank==0) and (problem.hall_of_fame_flag):
+            my_population.setup_hall_of_fame()
 
         if len(problem.genome_seeds) == 0:
             genome_seeds = []

codes/population.py

@@ -30,6 +30,7 @@ def __init__(self):
         #self.pop_size = population_size #moved to problem.pop_size
         self.population = []
         self.hall_of_fame = None
+        self.hof_ids = []
 
 
     def __setitem__(self, node_index, value):
@@ -93,9 +94,13 @@ def merge_subpopulations(self, subpops):
         ezLogging.info("Combined %i sub populations into a single population" % (len(subpops)))
 
 
-    def setup_hall_of_fame(self, maxsize):
+    def setup_hall_of_fame(self):
         '''
-        https://deap.readthedocs.io/en/master/api/tools.html#deap.tools.HallOfFame
+        https://deap.readthedocs.io/en/master/api/tools.html#deap.tools.ParetoFront
+        https://github.com/DEAP/deap/blob/master/deap/tools/support.py#L591
+        NOTE using ParetoFront which inherits from HallOfFame. See Issue #285.
+        But tldr is that HallOfFame only works on single objective and ParetoFront
+        works for multiobjective and no maxsize which should also addess Issue #280.
 
         letting hall_of_fame use to be optional
         '''
@@ -110,12 +115,13 @@ def similarity_equation(a, b):
             '''
             if a.id == b.id:
                 return True
+            elif a.fitness.values == b.fitness_values:
+                return True
             else:
                 return False
 
-        self.hall_of_fame = deap.tools.HallOfFame(maxsize=maxsize,
-                                                  similar=similarity_equation)
-        ezLogging.debug("Established 'Hall of Fame' with maxsize %i" % maxsize)
+        self.hall_of_fame = deap.tools.ParetoFront(similar=similarity_equation)
+        ezLogging.debug("Established Hall of Fame")
 
 
     def update_hall_of_fame(self):
@@ -127,10 +133,17 @@ def update_hall_of_fame(self):
             alive_population = []
             for indiv in self.population:
                 if not indiv.dead:
-                    alive_population.append(indiv)
+                    if indiv.id not in self.hof_ids:
+                        alive_population.append(indiv)
+
             self.hall_of_fame.update(alive_population)
             ezLogging.debug("Updated Hall of Fame to size %i" % (len(self.hall_of_fame.items)))
 
+        # keep a running list of HOF id's just to keep handy...won't be too much extra space
+        self.hof_ids = []
+        for indiv in self.hall_of_fame.items:
+            self.hof_ids.append(indiv.id)
+
 
     def get_pareto_front(self, use_hall_of_fame=False, first_front_only=False):
         '''

codes/utilities/simgan_feature_eval.py

@@ -8,6 +8,16 @@
 import torch
 from scipy.stats import ks_2samp, wasserstein_distance, ttest_ind
 
+
+### sys relative to root dir
+import sys
+from os.path import dirname, realpath
+sys.path.append(dirname(dirname(dirname(realpath(__file__)))))
+
+### absolute imports wrt root
+from codes.utilities.custom_logging import ezLogging
+
+
 def calc_feature_distances(refiners, validation_data, device):
     '''
     TODO...get the source from the room
@@ -726,26 +736,38 @@ def __init__(self, use_phdif=True, use_phratio=True, use_pdist=True, use_thdiff=
         self.use_changes = use_changes
         self.use_roughness = use_roughness
         self.use_th = use_th
+        self.feature_names = []
 
         self.num_features = 0
         if use_phdif:
             self.num_features += 1
+            self.feature_names.append("phdif")
         if use_phratio:
             self.num_features += 1
+            self.feature_names.append("phratio")
         if use_pdist:
             self.num_features += 1
+            self.feature_names.append("pdist")
         if use_thdiff:
             self.num_features += 2
+            self.feature_names.append("thdiff0")
+            self.feature_names.append("thdiff1")
         if use_thdist:
             self.num_features += 2
+            self.feature_names.append("thdist0")
+            self.feature_names.append("thdist1")
         if use_area:
             self.num_features += 1
+            self.feature_names.append("area")
         if use_changes:
             self.num_features += 1
+            self.feature_names.append("changes")
         if use_roughness:
             self.num_features += 1
+            self.feature_names.append("roughness")
         if use_th:
             self.num_features += 1
+            self.feature_names.append("th")
 
     def get_features(self, dataset):
         '''
@@ -755,6 +777,8 @@ def get_features(self, dataset):
 
             Returns:
                 features (ndarray): MxN' numpy array of N' features for M signals
+                (Rodd Added) - LIES! it's actually N'xM so we have to transpose whatever gets returned
+                    ...could be fixed if np.expand_dims axis is changed from 0 to 1 I think
         '''
         features = []
         if self.use_phdif:
@@ -823,7 +847,7 @@ def get_peak_distance(data, lrs=10, lre=40, rrs=40, rre=80):
             distance (ndarray): numpy array of the distances between the left and right peaks
     '''
     left_peak_inds = get_peaks(data, lrs, lre)
-    right_peak_inds = get_peaks(data, rrs, rre) 
+    right_peak_inds = get_peaks(data, rrs, rre)
     distances = right_peak_inds - left_peak_inds
     return distances