Non-multilevel Optimizers reworked (#507)

* optimizers passing with the new DataObjects except multilevel

developer_tools/XSDSchemas/Optimizers.xsd

@@ -16,6 +16,7 @@
             <xsd:element name="initialSeed" type="xsd:integer"       minOccurs="0"/>
             <xsd:element name="type"        type="xsd:string"        minOccurs="0"/>
             <xsd:element name="thresholdTrajRemoval" type="xsd:float" minOccurs="0"/>
+            <xsd:element name="writeSteps" type="xsd:string" minOccurs="0"/>
         </xsd:all>
     </xsd:complexType>
 

doc/user_manual/optimizer.tex

@@ -197,6 +197,12 @@ \subsubsection{Simultaneous Perturbation Stochastic Approximation (SPSA)}
     of the second.  Note that this value is
     calculated as Euclidean distance in a normalized 0 to 1 cubic domain, not the original input space domain.
     \default{0.05}
+    \item \xmlNode{writeSteps},  \xmlDesc{string, optional field}, specifies how often the current optimal
+      point should be stored to the solution export.  Options are \xmlString{every}, in which case each new
+      optimal point will be stored in the solution export; or \xmlString{final}, in which case only the most
+      optimal point found during the simulation will be stored in the solution export.
+    \default{every};
+
   \end{itemize}
 \end{itemize}
 \begin{itemize}
@@ -429,8 +435,8 @@ \subsubsection{Simultaneous Perturbation Stochastic Approximation (SPSA)}
 \subsubsection{Finite Difference Gradient Optimizer (FiniteDifferenceGradientOptimizer)}
 \label{subsubsubsec:FiniteDifferenceGradientOptimizer}
 The \textbf{FiniteDifferenceGradientOptimizer} optimization approach is  the simplest Gradient based approach since it is based on the
-first order evaluation of the Gradient.  A minimal number of $n variable$ 
-model evaluations are required in order to get a first order approximation of the gradient.  
+first order evaluation of the Gradient.  A minimal number of $n variable$
+model evaluations are required in order to get a first order approximation of the gradient.
 
 Current implementation of \textbf{FiniteDifferenceGradientOptimizer} can also handles
 constrained optimization problem. This paragraph briefly describes how current implementation ensures the input satisfies the
@@ -645,7 +651,7 @@ \subsubsection{Finite Difference Gradient Optimizer (FiniteDifferenceGradientOpt
         \cite{spall1998implementation}. A practical suggestion for $\gamma$ is 0.101 (paired with an
         $\alpha$ value of 0.602); however, the asymptotic limit is $\gamma=1/6$ ($\alpha=1$). \default{0.101}
       \item \xmlNode{c}, \xmlDesc{float, optional field} Step size coefficient.  This term determines the
-        nominal step size, and increasing it will directly increase the distance between points sampled in evaluating the gradient. 
+        nominal step size, and increasing it will directly increase the distance between points sampled in evaluating the gradient.
         It is suggested this parameter be approximately equal to the standard
         deviation of the measurement noise in the response for stochastic responses.  For regular responses,
         it can be a small arbitrary value. \default{0.005}
@@ -706,4 +712,4 @@ \subsubsection{Finite Difference Gradient Optimizer (FiniteDifferenceGradientOpt
   </FiniteDifferenceGradientOptimizer>
   ...
 </Optimizers>
-\end{lstlisting}
+\end{lstlisting}

framework/Optimizers/GradientBasedOptimizer.py

@@ -296,7 +296,7 @@ def _updateConvergenceVector(self, traj, varsUpdate, currentLossVal):
 
     ## first, determine if we want to keep the new point
     # obtain the old loss value
-    oldLossVal = self.counter['recentOptHist'][traj][0]['output']
+    oldLossVal = self.counter['recentOptHist'][traj][0][self.objVar]
     # see if new point is better than old point
     newerIsBetter = self.checkIfBetter(currentLossVal,oldLossVal)
     # if this was a recommended preconditioning point, we should not be converged.
@@ -379,7 +379,7 @@ def printProgress(name,boolCheck,test,gold):
 
         #same coordinate check
         oldInputSpace = set(self.optVarsHist[traj][varsUpdate].items())
-        curInputSpace = set(self.counter['recentOptHist'][traj][0]['inputs'].items())
+        curInputSpace = set(dict((var,self.counter['recentOptHist'][traj][0][var]) for var in self.getOptVars()).items())
         sameCoordinateCheck = oldInputSpace == curInputSpace
         self.raiseAMessage(printString.format('Same coordinate check',str(minStepSizeCheck)))
         converged = converged or sameCoordinateCheck
@@ -548,65 +548,80 @@ def localFinalizeActualSampling(self,jobObject,model,myInput):
               except KeyError:
                 # this means we don't have an entry for this trajectory yet, so don't copy anything
                 pass
-              self.counter['recentOptHist'][traj][0] = {}
-              self.counter['recentOptHist'][traj][0]['inputs'] = self.optVarsHist[traj][self.counter['varsUpdate'][traj]]
-              self.counter['recentOptHist'][traj][0]['output'] = currentObjectiveValue
+              # store realization of most recent developments
+              rlz = {}
+              rlz.update(self.optVarsHist[traj][self.counter['varsUpdate'][traj]])
+              rlz.update(outputs)
+              self.counter['recentOptHist'][traj][0] = rlz
               if traj not in self.counter['prefixHistory']:
                 self.counter['prefixHistory'][traj] = []
               self.counter['prefixHistory'][traj].append(prefix)
             # update solution export
             #FIXME much of this should move to the base class!
             if not failedTraj:
-              # create realization to add to data object
-              rlz = {}
-              badValue = -1 #value to use if we don't have a value # TODO make this accessible to user?
-              for var in self.solutionExport.getVars():
-                if var in self.getOptVars():
-                  new = self.denormalizeData(self.counter['recentOptHist'][traj][0]['inputs'])[var]
-                elif var == self.objVar:
-                  new = self.counter['recentOptHist'][traj][0]['output']
-                elif var in outputs.keys():
-                  new = outputs[var]
-                elif var == 'varsUpdate':
-                  new = self.counter['solutionUpdate'][traj]
-                elif var == 'trajID':
-                  new = traj+1 # +1 is for historical reasons, when histories were indexed on 1 instead of 0
-                elif var == 'stepSize':
-                  try:
-                    new = self.counter['lastStepSize'][traj]
-                  except KeyError:
-                    new = badValue
-                elif var.startswith( 'gradient_'):
-                  varName = var[9:]
-                  vec = self.counter['gradientHistory'][traj][0].get(varName,None)
-                  if vec is not None:
-                    new = vec*self.counter['gradNormHistory'][traj][0]
-                  else:
-                    new = badValue
-                elif var.startswith( 'convergenceAbs'):
-                  try:
-                    new = self.convergenceProgress[traj].get('abs',badValue)
-                  except KeyError:
-                    new = badValue
-                elif var.startswith( 'convergenceRel'):
-                  try:
-                    new = self.convergenceProgress[traj].get('rel',badValue)
-                  except KeyError:
-                    new = badValue
-                elif var.startswith( 'convergenceGrad'):
-                  try:
-                    new = self.convergenceProgress[traj].get('grad',badValue)
-                  except KeyError:
-                    new = badValue
-                else:
-                  self.raiseAnError(IOError,'Unrecognized output request:',var)
-                # format for realization
-                rlz[var] = np.atleast_1d(new)
-              self.solutionExport.addRealization(rlz)
+              # only write here if we want to write on EVERY optimizer iteration (each new optimal point)
+              if self.writeSolnExportOn == 'every':
+                self.writeToSolutionExport(traj)
+              # whether we wrote to solution export or not, update the counter
               self.counter['solutionUpdate'][traj] += 1
           else: #not ready to update solutionExport
             break
 
+  def writeToSolutionExport(self,traj):
+    """
+      Standardizes how the solution export is written to.
+      Uses data from "recentOptHist" and other counters to fill in values.
+      @ In, traj, int, the trajectory for which an entry is being written
+      @ Out, None
+    """
+    # create realization to add to data object
+    rlz = {}
+    badValue = -1 #value to use if we don't have a value # TODO make this accessible to user?
+    recent = self.counter['recentOptHist'][traj][0]
+    for var in self.solutionExport.getVars():
+      # inputs, objVar, other outputs
+      if var in recent.keys():
+        new = self.denormalizeData(recent)[var]
+      # custom counters: varsUpdate, trajID, stepSize
+      elif var == 'varsUpdate':
+        new = self.counter['solutionUpdate'][traj]
+      elif var == 'trajID':
+        new = traj+1 # +1 is for historical reasons, when histories were indexed on 1 instead of 0
+      elif var == 'stepSize':
+        try:
+          new = self.counter['lastStepSize'][traj]
+        except KeyError:
+          new = badValue
+      # variable-dependent information: gradients
+      elif var.startswith( 'gradient_'):
+        varName = var[9:]
+        vec = self.counter['gradientHistory'][traj][0].get(varName,None)
+        if vec is not None:
+          new = vec*self.counter['gradNormHistory'][traj][0]
+        else:
+          new = badValue
+      # convergence metrics
+      elif var.startswith( 'convergenceAbs'):
+        try:
+          new = self.convergenceProgress[traj].get('abs',badValue)
+        except KeyError:
+          new = badValue
+      elif var.startswith( 'convergenceRel'):
+        try:
+          new = self.convergenceProgress[traj].get('rel',badValue)
+        except KeyError:
+          new = badValue
+      elif var.startswith( 'convergenceGrad'):
+        try:
+          new = self.convergenceProgress[traj].get('grad',badValue)
+        except KeyError:
+          new = badValue
+      else:
+        self.raiseAnError(IOError,'Unrecognized output request:',var)
+      # format for realization
+      rlz[var] = np.atleast_1d(new)
+    self.solutionExport.addRealization(rlz)
+
   def fractionalStepChangeFromGradHistory(self,traj):
     """
       Uses the dot product between two successive gradients to determine a fractional multiplier for the step size.
@@ -687,3 +702,27 @@ def proposeNewPoint(self,traj,point):
     Optimizer.proposeNewPoint(self,traj,point)
     self.counter['varsUpdate'][traj] += 1 #usually done when evaluating gradient, but we're bypassing that
     self.queueUpOptPointRuns(traj,self.recommendedOptPoint[traj])
+
+  def finalizeSampler(self,failedRuns):
+    """
+      Method called at the end of the Step when no more samples will be taken.  Closes out optimizer.
+      @ In, failedRuns, list, list of JobHandler.ExternalRunner objects
+      @ Out, None
+    """
+    Optimizer.handleFailedRuns(self,failedRuns)
+    # if writing soln export only on final, now is the time to do it
+    if self.writeSolnExportOn == 'final':
+      # get the most optimal point among the trajectories
+      bestValue = None
+      bestTraj = None
+      for traj in self.counter['recentOptHist'].keys():
+        value = self.counter['recentOptHist'][traj][0][self.objVar]
+        if bestTraj is None:
+          bestTraj = traj
+          bestValue = value
+          continue
+        if self.checkIfBetter(value,bestValue):
+          bestTraj = traj
+          bestValue = value
+      # now have the best trajectory, so write solution export
+      self.writeToSolutionExport(bestTraj)

framework/Optimizers/Optimizer.py

@@ -81,6 +81,7 @@ def __init__(self):
     self.optVarsInit['ranges']          = {}                        # Dict of the ranges (min and max) of each variable's domain
     self.initSeed                       = None                      # Seed for random number generators
     self.optType                        = None                      # Either max or min
+    self.writeSolnExportOn              = None                      # Determines when we write to solution export (every step or final solution)
     self.paramDict                      = {}                        # Dict containing additional parameters for derived class
     self.initializationSampler          = None                      # Sampler that can be used to initialize the optimizer trajectories
     self.optVarsInitialized             = {}                        # Dict {var1:<initial> present?,var2:<initial> present?}
@@ -249,8 +250,15 @@ def _readMoreXMLbase(self,xmlNode):
             self.initSeed = int(childChild.text)
           elif childChild.tag == 'thresholdTrajRemoval':
             self.thresholdTrajRemoval = float(childChild.text)
+          elif childChild.tag == 'writeSteps':
+            if childChild.text.strip().lower() == 'every':
+              self.writeSolnExportOn = 'every'
+            elif childChild.text.strip().lower() == 'final':
+              self.writeSolnExportOn = 'final'
+            else:
+              self.raiseAnError(IOError,'Unexpected frequency for <writeSteps>: "{}". Expected "every" or "final".')
           else:
-            self.raiseAnError(IOError,'Unknown tag '+childChild.tag+' .Available: limit, type, initialSeed!')
+            self.raiseAnError(IOError,'Unknown tag: '+childChild.tag)
 
       elif child.tag == "convergence":
         for childChild in child:
@@ -297,6 +305,10 @@ def _readMoreXMLbase(self,xmlNode):
           elif subnode.tag == 'sequence':
             self.mlSequence = list(x.strip() for x in subnode.text.split(','))
 
+    if self.writeSolnExportOn is None:
+      # default
+      self.writeSolnExportOn = 'every'
+    self.raiseAMessage('Writing to solution export on "{}" optimizer iteration.'.format(self.writeSolnExportOn))
     if self.optType is None:
       self.optType = 'min'
     if self.thresholdTrajRemoval is None:
@@ -660,7 +672,7 @@ def amIreadyToProvideAnInput(self):
         # do we have any opt points yet?
         if len(self.counter['recentOptHist'][traj][0]) > 0:
           # get the latset optimization point (normalized)
-          latestPoint = self.counter['recentOptHist'][traj][0]['inputs']
+          latestPoint = dict((var,self.counter['recentOptHist'][traj][0][var]) for var in self.getOptVars())
           #some flags for clarity of checking
           justStarted = self.mlDepth[traj] is None
           inInnermost = self.mlDepth[traj] is not None and self.mlDepth[traj] == len(self.mlSequence)-1
@@ -996,6 +1008,14 @@ def _getJobsByID(self):
     """
     pass
 
+  def finalizeSampler(self,failedRuns):
+    """
+      Method called at the end of the Step when no more samples will be taken.  Closes out the optimizer for this step.
+      @ In, failedRuns, list, list of JobHandler.ExternalRunner objects
+      @ Out, None
+    """
+    self.handleFailedRuns(failedRuns)
+
   def handleFailedRuns(self,failedRuns):
     """
       Collects the failed runs from the Step and allows optimizer to handle them individually if need be.

framework/Optimizers/SPSA.py

@@ -119,7 +119,7 @@ def _newOptPointAdd(self, gradient, traj):
     """
     ak = self._computeGainSequenceAk(self.paramDict,self.counter['varsUpdate'][traj],traj) # Compute the new ak
     self.optVarsHist[traj][self.counter['varsUpdate'][traj]] = {}
-    varK = copy.deepcopy(self.counter['recentOptHist'][traj][0]['inputs'])
+    varK = dict((var,self.counter['recentOptHist'][traj][0][var]) for var in self.getOptVars(traj)) #copy.deepcopy(self.counter['recentOptHist'][traj][0]['inputs'])
     varKPlus,modded = self._generateVarsUpdateConstrained(traj,ak,gradient,varK)
     #check for redundant paths
     if len(self.optTrajLive) > 1 and self.counter['solutionUpdate'][traj] > 0:
@@ -330,7 +330,7 @@ def localGenerateInput(self,model,oldInput):
       if self.counter['perturbation'][traj] == 1:
         # Generate all the perturbations at once, then we can submit them one at a time
         ck = self._computeGainSequenceCk(self.paramDict,self.counter['varsUpdate'][traj]+1)
-        varK = self.counter['recentOptHist'][traj][0]['inputs']
+        varK = dict((var,self.counter['recentOptHist'][traj][0][var]) for var in self.getOptVars(traj))
         #check the submission queue is empty; otherwise something went wrong # TODO this is a sanity check, might be removed for efficiency
         #TODO this same check is in GradientBasedOptimizer.queueUpOptPointRuns, they might benefit from abstracting
         if len(self.submissionQueue[traj]) > 0:

framework/PostProcessorFunctions/HistorySetSnapShot.py

@@ -199,17 +199,32 @@ def historySnapShot(inputDic, pivotVar, snapShotType, pivotVal=None, tempID = No
     @ In, tempID, string, name of the temporal variable (default None)
     @ Out, outputDic, dict, it contains the temporal slice of all histories
   """
+  # place to store data results
   outputDic={'data':{}}
-  outputDic['data']['ProbabilityWeight'] = inputDic['data']['ProbabilityWeight']
-  outputDic['data']['prefix'] = inputDic['data']['prefix']
+  # collect metadata, if it exists, to pass through
+  # TODO collecting by name is problemsome; for instance, Optimizers don't produce "probability weight" information
+  ## ProbabilityWeight
+  try:
+    outputDic['data']['ProbabilityWeight'] = inputDic['data']['ProbabilityWeight']
+  except KeyError:
+    pass
+  ## prefix
+  try:
+    outputDic['data']['prefix'] = inputDic['data']['prefix']
+  except KeyError:
+    pass
+  # place to store dimensionalities
   outputDic['dims'] = {key: [] for key in inputDic['dims'].keys()}
 
   for var in inputDic['inpVars']:
     outputDic['data'][var] = inputDic['data'][var]
 
   outVars = inputDic['data'].keys()
   outVars = [var for var in outVars if 'Probability' not in var]
-  outVars.remove('prefix')
+  try:
+    outVars.remove('prefix')
+  except ValueError:
+    pass
   vars = [var for var in outVars if var not in inputDic['inpVars']]
 
   for var in vars:

framework/Samplers/Sampler.py

@@ -826,6 +826,14 @@ def _reassignPbWeightToCorrelatedVars(self):
         for subVarName in varName.split(","):
           self.inputInfo['ProbabilityWeight-' + subVarName.strip()] = self.inputInfo['ProbabilityWeight-' + varName]
 
+  def finalizeSampler(self,failedRuns):
+    """
+      Method called at the end of the Step when no more samples will be taken.  Closes out sampler for step.
+      @ In, failedRuns, list, list of JobHandler.ExternalRunner objects
+      @ Out, None
+    """
+    self.handleFailedRuns(failedRuns)
+
   def handleFailedRuns(self,failedRuns):
     """
       Collects the failed runs from the Step and allows samples to handle them individually if need be.

framework/Steps.py

@@ -687,8 +687,9 @@ def _localTakeAstepRun(self,inDictionary):
         self.raiseADebug('Finished with %d runs submitted, %d jobs running, and %d completed jobs waiting to be processed.' % (jobHandler.numSubmitted(),jobHandler.numRunning(),len(jobHandler.getFinishedNoPop())) )
         break
       time.sleep(self.sleepTime)
-    # if any new collected runs failed, let the sampler treat them appropriately
-    sampler.handleFailedRuns(self.failedRuns)
+    # END while loop that runs the step iterations
+    # if any collected runs failed, let the sampler treat them appropriately, and any other closing-out actions
+    sampler.finalizeSampler(self.failedRuns)
 
   def _findANewInputToRun(self, sampler, model, inputs, outputs):
     """