tweaks

nifty/energies/quadratic_energy.py

@@ -8,14 +8,22 @@ class QuadraticEnergy(Energy):
     position-independent.
     """
 
-    def __init__(self, position, A, b):
+    def __init__(self, position, A, b, grad=None):
         super(QuadraticEnergy, self).__init__(position=position)
         self._A = A
         self._b = b
+        if grad is not None:
+            self._Ax = grad + self._b
+        else:
+            self._Ax = self._A(self.position)
 
     def at(self, position):
         return self.__class__(position=position, A=self._A, b=self._b)
 
+    def at_with_grad(self, position, grad):
+        return self.__class__(position=position, A=self._A, b=self._b,
+                              grad=grad)
+
     @property
     @memo
     def value(self):
@@ -29,8 +37,3 @@ class QuadraticEnergy(Energy):
     @property
     def curvature(self):
         return self._A
-
-    @property
-    @memo
-    def _Ax(self):
-        return self.curvature(self.position)

nifty/minimization/conjugate_gradient.py

@@ -79,7 +79,7 @@ class ConjugateGradient(Minimizer):
         if self._preconditioner is not None:
             d = self._preconditioner(r)
         else:
-            d = r
+            d = r.copy()
         previous_gamma = (r.vdot(d)).real
         if previous_gamma == 0:
             return energy, controller.CONVERGED
@@ -92,11 +92,6 @@ class ConjugateGradient(Minimizer):
                 return energy, controller.ERROR
             alpha = previous_gamma/ddotq
 
-            energy = energy.at(energy.position+d*alpha)
-            status = self._controller.check(energy)
-            if status != controller.CONTINUE:
-                return energy, status
-
             reset = False
             if alpha < 0:
                 self.logger.warn("Positive definiteness of A violated!")
@@ -105,9 +100,15 @@ class ConjugateGradient(Minimizer):
                 reset += (iteration_number % self._reset_count == 0)
             if reset:
                 self.logger.info("Resetting conjugate directions.")
+                energy = energy.at(energy.position+d*alpha)
                 r = -energy.gradient
             else:
                 r -= q * alpha
+                energy = energy.at_with_grad(energy.position+d*alpha,-r)
+
+            status = self._controller.check(energy)
+            if status != controller.CONTINUE:
+                return energy, status
 
             if self._preconditioner is not None:
                 s = self._preconditioner(r)

nifty/minimization/default_iteration_controller.py

@@ -19,30 +19,34 @@
 from .iteration_controller import IterationController
 
 class DefaultIterationController(IterationController):
-    def __init__ (self, tol_gradnorm=None, convergence_level=1,
-                  iteration_limit=None):
+    def __init__ (self, tol_gradnorm=None, tol_rel_gradnorm=None,
+                  convergence_level=1, iteration_limit=None):
         super(DefaultIterationController, self).__init__()
         self._tol_gradnorm = tol_gradnorm
+        self._tol_rel_gradnorm = tol_rel_gradnorm
         self._convergence_level = convergence_level
         self._iteration_limit = iteration_limit
 
     def start(self, energy):
         self._itcount = -1
         self._ccount = 0
+        if self._tol_rel_gradnorm is not None:
+            self._tol_rel_gradnorm *= energy.gradient_norm
         return self.check(energy)
 
     def check(self, energy):
         self._itcount += 1
-        print "iteration",self._itcount,"gradnorm",energy.gradient_norm,"level",self._ccount
+        print "iteration",self._itcount,"gradnorm",energy.gradient_norm,"level",self._ccount, self._tol_rel_gradnorm
         if self._iteration_limit is not None:
             if self._itcount >= self._iteration_limit:
                 return self.CONVERGED
         if self._tol_gradnorm is not None:
             if energy.gradient_norm <= self._tol_gradnorm:
                 self._ccount += 1
-            if self._ccount >= self._convergence_level:
-                return self.CONVERGED
-        else:
-            self._ccount = max(0, self._ccount-1)
+        if self._tol_rel_gradnorm is not None:
+            if energy.gradient_norm <= self._tol_rel_gradnorm:
+                self._ccount += 1
+        if self._ccount >= self._convergence_level:
+            return self.CONVERGED
 
         return self.CONTINUE

nifty/minimization/descent_minimizer.py

@@ -75,7 +75,7 @@ class DescentMinimizer(Minimizer):
         controller = self._controller
         status = controller.start(energy)
         if status != controller.CONTINUE:
-            return E, status
+            return energy, status
 
         while True:
             # check if position is at a flat point

nifty/minimization/relaxed_newton.py

@@ -22,9 +22,8 @@ from .line_searching import LineSearchStrongWolfe
 
 class RelaxedNewton(DescentMinimizer):
     def __init__(self, controller, line_searcher=LineSearchStrongWolfe()):
-        super(RelaxedNewton, self).__init__(
-                                controller=controller,
-                                line_searcher=line_searcher)
+        super(RelaxedNewton, self).__init__(controller=controller,
+                                            line_searcher=line_searcher)
 
         self.line_searcher.preferred_initial_step_size = 1.