- Bugfixes and improvements on ProbabilityDistribution and KDE, msfts

- Scale Transformation

benchmarks/Measures.py

@@ -234,7 +234,7 @@ def get_point_statistics(data, model, indexer=None):
     """Condensate all measures for point forecasters"""
 
     if indexer is not None:
-        ndata = np.array(indexer.get_data(data[model.order:]))
+        ndata = np.array(indexer.get_data(data))
     else:
         ndata = np.array(data[model.order:])
 

common/FLR.py

@@ -62,7 +62,7 @@ def generateRecurrentFLRs(fuzzyData):
     return flrs
 
 
-def generateIndexedFLRs(sets, indexer, data):
+def generateIndexedFLRs(sets, indexer, data, transformation=None):
     """
     Create a season-indexed ordered FLR set from a list of fuzzy sets with recurrence
     :param sets: fuzzy sets
@@ -73,7 +73,9 @@ def generateIndexedFLRs(sets, indexer, data):
     flrs = []
     index = indexer.get_season_of_data(data)
     ndata = indexer.get_data(data)
-    for k in np.arange(1,len(data)):
+    if transformation is not None:
+        ndata = transformation.apply(ndata)
+    for k in np.arange(1,len(ndata)):
         lhs = FuzzySet.getMaxMembershipFuzzySet(ndata[k-1],sets)
         rhs = FuzzySet.getMaxMembershipFuzzySet(ndata[k], sets)
         season = index[k]

common/Transformations.py

@@ -73,6 +73,41 @@ class Differential(Transformation):
             return inc
 
 
+class Scale(Transformation):
+    def __init__(self, min=0, max=1):
+        super(Scale, self).__init__([min, max])
+        self.data_max = None
+        self.data_min = None
+        self.transf_max = max
+        self.transf_min = min
+
+    def apply(self, data, param=None,**kwargs):
+        if self.data_max is None:
+            self.data_max = np.nanmax(data)
+            self.data_min = np.nanmin(data)
+        data_range = self.data_max - self.data_min
+        transf_range = self.transf_max - self.transf_min
+        if isinstance(data, list):
+            tmp = [(k + (-1 * self.data_min)) / data_range for k in data]
+            tmp2 = [ (k * transf_range) + self.transf_min for k in tmp]
+        else:
+            tmp = (data + (-1 * self.data_min)) / data_range
+            tmp2 = (tmp * transf_range) + self.transf_min
+
+        return  tmp2
+
+    def inverse(self, data, param, **kwargs):
+        data_range = self.data_max - self.data_min
+        transf_range = self.transf_max - self.transf_min
+        if isinstance(data, list):
+            tmp2 = [(k - self.transf_min) / transf_range   for k in data]
+            tmp = [(k * data_range) + self.data_min for k in tmp2]
+        else:
+            tmp2 = (data - self.transf_min) / transf_range
+            tmp = (tmp2 * data_range) + self.data_min
+        return tmp
+
+
 class AdaptiveExpectation(Transformation):
     """
     Adaptive Expectation post processing

ensemble/multiseasonal.py

@@ -82,7 +82,8 @@ class SeasonalEnsembleFTS(ensemble.EnsembleFTS):
 
             tmp = np.ravel(tmp).tolist()
 
-            dist = ProbabilityDistribution.ProbabilityDistribution("KDE",h=h,uod=[self.original_min, self.original_max], data=tmp)
+            dist = ProbabilityDistribution.ProbabilityDistribution("KDE", uod=[self.original_min, self.original_max], data=tmp,
+                                                                   **kwargs)
 
             ret.append(dist)
 

models/msfts.py

@@ -13,10 +13,10 @@ class MultiSeasonalFTS(sfts.SeasonalFTS):
         self.shortname = "MSFTS " + name
         self.detail = ""
         self.seasonality = 1
-        self.hasSeasonality = True
+        self.has_seasonality = True
-        self.hasPointForecasting = True
+        self.has_point_forecasting = True
-        self.isHighOrder = True
+        self.is_high_order = False
-        self.isMultivariate = True
+        self.is_multivariate = True
         self.indexer = indexer
         self.flrgs = {}
 

probabilistic/ProbabilityDistribution.py

@@ -123,8 +123,11 @@ class ProbabilityDistribution(object):
             ys = [self.distribution[k]/self.count for k in self.bins]
         else:
             ys = [self.distribution[k] for k in self.bins]
+            yp = [0 for k in self.data]
+            axis.plot(self.data, yp, c="red")
 
-        axis.plot(self.bins, ys,c=color, label=self.name)
+
+        axis.plot(self.bins, ys, c=color, label=self.name)
 
         axis.set_xlabel('Universe of Discourse')
         axis.set_ylabel('Probability')
@@ -134,5 +137,8 @@ class ProbabilityDistribution(object):
         body = '|'
         for k in sorted(self.distribution.keys()):
             head += str(round(k,2)) + '\t|'
+            if self.type == "histogram":
                 body += str(round(self.distribution[k]  / self.count,3)) + '\t|'
+            else:
+                body += str(round(self.distribution[k], 3)) + '\t|'
         return head + '\n' + body

probabilistic/kde.py

@@ -1,3 +1,6 @@
+from pyFTS.common import Transformations
+import numpy as np
+
 """
 Kernel Density Estimation 
 """
@@ -8,17 +11,21 @@ class KernelSmoothing(object):
     def __init__(self,h, method="epanechnikov"):
         self.h = h
         self.method = method
+        self.transf = Transformations.Scale(min=0,max=1)
 
     def kernel(self, u):
         if self.method == "epanechnikov":
-            return (3/4) * (1 - u**2)
+            return (3/4)*(1 - u**2)
         elif self.method == "gaussian":
-            return 0.5
+            return (1/np.sqrt(2*np.pi))*np.exp(-0.5*u**2)
         elif self.method == "uniform":
             return 0.5
 
     def probability(self, x, data):
         l = len(data)
-        p = sum([self.kernel((x - k)/self.h) for k in data]) / l*self.h
+
+        ndata = self.transf.apply(data)
+        nx = self.transf.apply(x)
+        p = sum([self.kernel((nx - k)/self.h) for k in ndata]) / l*self.h
 
         return p

tests/general.py

@@ -62,22 +62,43 @@ DATASETS
 
 #print(lag)
 #print(a)
-
+#'''
-sonda = pd.read_csv("DataSets/SONDA_BSB_MOD.csv", sep=";")
+sonda = pd.read_csv("DataSets/SONDA_BSB_15MIN_AVG.csv", sep=";")
 
 sonda['data'] = pd.to_datetime(sonda['data'])
 
-sonda = sonda[:][527041:]
+#sonda = sonda[:][527041:].dropna()
 
 sonda.index = np.arange(0,len(sonda.index))
 
-sonda_treino = sonda[:1051200]
+sonda_treino = sonda[:105313].dropna()
-sonda_teste = sonda[1051901:1051910]
+sonda_teste = sonda[105314:].dropna()
 
-ix_m15 = SeasonalIndexer.DateTimeSeasonalIndexer('data',[SeasonalIndexer.DateTime.minute],[15],'glo_avg', name='m15')
+#ix_m15 = SeasonalIndexer.DateTimeSeasonalIndexer('data',[SeasonalIndexer.DateTime.minute],[15],'glo_avg', name='m15')
 
-fs1 = Grid.GridPartitioner(sonda_treino,50,transformation=diff, indexer=ix_m15)
+#fs1 = Grid.GridPartitioner(sonda_treino, 50, transformation=diff, indexer=ix_m15)
 
+#ix = cUtil.load_obj("models/sonda_ix_Mhm15.pkl")
+
+#fs = cUtil.load_obj("models/sonda_fs_Entropy40_diff.pkl")
+
+#from pyFTS.models import msfts
+
+#obj = msfts.MultiSeasonalFTS("sonda_msfts_Entropy40_Mhm15", indexer=ix)
+
+#obj.appendTransformation(diff)
+
+#obj.train(sonda_treino, fs.sets)
+
+#cUtil.persist_obj(obj, "models/sonda_msfts_Entropy40_Mhm15.pkl")
+
+ftse = cUtil.load_obj("models/sonda_ensemble_msfts.pkl")
+
+tmp = ftse.forecastDistribution(sonda_teste[850:860],  h=0.5, method="gaussian")
+
+print(tmp[0])
+
+#'''
 
 '''
 from pyFTS.models.seasonal import SeasonalIndexer
@@ -223,7 +244,7 @@ bchmk.ahead_sliding_window(sonda, 10000, steps=10, resolution=10, train=0.2, inc
                      dump=True, save=True, file="experiments/sondawind_ahead_analytic_diff.csv",
                      nodes=['192.168.0.106', '192.168.0.108', '192.168.0.109']) #, depends=[hofts, ifts])
 
-"""
+
 
 from pyFTS import pwfts
 from pyFTS.common import Transformations
@@ -240,7 +261,7 @@ from pyFTS.partitioners import Grid
 #bchmk.plot_compared_intervals_ahead(best[1600:1700],[model], ['blue','red'],
 #                                    distributions=[True], save=True, file="pictures/best_ahead_forecasts",
 #                                    time_from=40, time_to=60, resolution=100)
-'''
+
 experiments = [
     ["experiments/taiex_point_synthetic_diff.csv","experiments/taiex_point_analytic_diff.csv",16],
     ["experiments/nasdaq_point_synthetic_diff.csv","experiments/nasdaq_point_analytic_diff.csv", 11],
@@ -320,7 +341,6 @@ diff = Transformations.Differential(1)
 fs = Grid.GridPartitioner(sonda[:9000], 10, transformation=diff)
 
 
-'''
 tmp = sfts.SeasonalFTS("")
 tmp.indexer = ix
 tmp.appendTransformation(diff)