Merge pull request #104 from antoinecarme/ts_cross_val

Add the possibility to use cross validation when training PyAF models #105
antoinecarme · Sep 27, 2018 · 264106f · 264106f
2 parents 362c968 + 6e7bacd
commit 264106f
Show file tree

Hide file tree

Showing 88 changed files with 16,270 additions and 3,250 deletions.
diff --git a/TS/Options.py b/TS/Options.py
@@ -93,6 +93,10 @@ def disable_all_periodics(self):
     def disable_all_autoregressions(self):
         self.set_active_autoregressions([]);
 
+class cCrossValidationOptions:
+    def __init__(self):
+        self.mMethod = None;
+        self.mNbFolds = 10
 
 
 class cSignalDecomposition_Options(cModelControl):
@@ -101,7 +105,8 @@ def __init__(self):
         super().__init__();
         self.mParallelMode = True;
         self.mNbCores = 8;
-        self.mEstimRatio = 0.8;
+        self.mEstimRatio = 0.8; # to be deprecated when cross validation is OK.
+        self.mCustomSplit = None
         self.mAddPredictionIntervals = True
         self.enable_fast_mode();
         self.mTimeDeltaComputationMethod = "AVG"; # can be "AVG", "MODE", "USER"
@@ -118,6 +123,7 @@ def __init__(self):
         self.mHierarchicalCombinationMethod = "BU";
         self.mForecastRectifier = None # can be "relu" to force positive forecast values
         self.mXGBOptions = None
+        self.mCrossValidationOptions = cCrossValidationOptions()
         self.disableDebuggingOptions();
 
     def disableDebuggingOptions(self):
@@ -142,6 +148,8 @@ def enable_slow_mode(self):
 
         self.mMaxAROrder = 64;
         self.mFilterSeasonals = False
+        # enable cross validation
+        self.mCrossValidationOptions.mMethod = "TSCV";        
 
     def enable_fast_mode(self):
         self.mQuantiles = [5, 10, 20]; # quintiles, deciles, and vingtiles;)

diff --git a/TS/PredictionIntervals.py b/TS/PredictionIntervals.py
@@ -31,20 +31,21 @@ def computePerformances(self):
         lForecastColumn = str(self.mSignal) + "_Forecast";
         df = self.mModel.mTrend.mSignalFrame.reset_index();
         N = df.shape[0];
-        (lOriginalFit, lOriginalForecast, lOriginalTest) = self.mModel.mTimeInfo.cutFrame(df);
+        (lOriginalFit, lOriginalForecast, lOriginalTest) = self.mModel.mTimeInfo.mSplit.cutFrame(df);
         df1 = df;
         for h in range(0 , self.mHorizon):
             df2 = None;
             df2 = self.mModel.forecastOneStepAhead(df1, perf_mode = True);
             df2 = df2.head(N);
             lHorizonName = lForecastColumn + "_" + str(h + 1);
-            (lFrameFit, lFrameForecast, lFrameTest) = self.mModel.mTimeInfo.cutFrame(df2);
+            (lFrameFit, lFrameForecast, lFrameTest) = self.mModel.mTimeInfo.mSplit.cutFrame(df2);
             self.mFitPerformances[lHorizonName] = tsperf.cPerf();
             self.mFitPerformances[lHorizonName].compute(lOriginalFit[lSignalColumn], lFrameFit[lForecastColumn], lHorizonName);
             self.mForecastPerformances[lHorizonName] = tsperf.cPerf();
             self.mForecastPerformances[lHorizonName].compute(lOriginalForecast[lSignalColumn], lFrameForecast[lForecastColumn], lHorizonName);
             self.mTestPerformances[lHorizonName] = tsperf.cPerf();
-            self.mTestPerformances[lHorizonName].compute(lOriginalTest[lSignalColumn], lFrameTest[lForecastColumn], lHorizonName);
+            if(lOriginalTest.shape[0] > 0):
+                self.mTestPerformances[lHorizonName].compute(lOriginalTest[lSignalColumn], lFrameTest[lForecastColumn], lHorizonName);
             df1 = df2[[lTimeColumn , lForecastColumn]];
             df1.columns = [lTimeColumn , lSignalColumn]
         # self.dump_detailed();

diff --git a/TS/Scikit_Models.py b/TS/Scikit_Models.py
@@ -32,7 +32,7 @@ def fit(self):
         series = self.mCycleResidueName; 
         self.mTime = self.mTimeInfo.mTime;
         self.mSignal = self.mTimeInfo.mSignal;
-        lAREstimFrame = self.mTimeInfo.getEstimPart(self.mARFrame)
+        lAREstimFrame = self.mSplit.getEstimPart(self.mARFrame)
 
         # print("mAREstimFrame columns :" , self.mAREstimFrame.columns);
         lARInputs = lAREstimFrame[self.mInputNames].values
@@ -154,10 +154,10 @@ def build_Scikit_Model(self):
 
     def set_name(self):
         self.mOutName = self.mCycleResidueName +  '_AR(' + str(self.mNbLags) + ")";
-        self.mFormula = "AR(" + str(self.mNbLags) + ")";
+        self.mFormula = "AR" # (" + str(self.mNbLags) + ")";
         if(self.mExogenousInfo is not None):
             self.mOutName = self.mCycleResidueName +  '_ARX(' + str(self.mNbLags) + ")";
-            self.mFormula = "ARX(" + str(self.mNbExogenousLags) + ")";
+            self.mFormula = "ARX" # (" + str(self.mNbExogenousLags) + ")";
 
 
 
@@ -176,10 +176,10 @@ def build_Scikit_Model(self):
 
     def set_name(self):
         self.mOutName = self.mCycleResidueName +  '_SVR(' + str(self.mNbLags) + ")";
-        self.mFormula = "SVR(" + str(self.mNbLags) + ")";
+        self.mFormula = "SVR" # (" + str(self.mNbLags) + ")";
         if(self.mExogenousInfo is not None):
             self.mOutName = self.mCycleResidueName +  '_SVRX(' + str(self.mNbLags) + ")";
-            self.mFormula = "SVRX(" + str(self.mNbExogenousLags) + ")";
+            self.mFormula = "SVRX" # (" + str(self.mNbExogenousLags) + ")";
 
 
 class cXGBoost_Model(cAbstract_Scikit_Model):
@@ -211,7 +211,7 @@ def build_Scikit_Model(self):
 
     def set_name(self):
         self.mOutName = self.mCycleResidueName +  '_XGB(' + str(self.mNbLags) + ")";
-        self.mFormula = "XGB(" + str(self.mNbLags) + ")";
+        self.mFormula = "XGB" #  + str(self.mNbLags) + ")";
         if(self.mExogenousInfo is not None):
             self.mOutName = self.mCycleResidueName +  '_XGBX(' + str(self.mNbLags) + ")";
-            self.mFormula = "XGBX(" + str(self.mNbExogenousLags) + ")";
+            self.mFormula = "XGBX" # (" + str(self.mNbExogenousLags) + ")";