Dolci/tape recompute count (#172)

* Add a tape computed counter.

Author: Ig-dolci

pyadjoint/reduced_functional.py

@@ -204,6 +204,7 @@ def __call__(self, values):
 
         self.tape.reset_blocks()
         blocks = self.tape.get_blocks()
+        self.tape._recompute_count += 1
         with self.marked_controls():
             with stop_annotating():
                 if self.tape._checkpoint_manager:

pyadjoint/tape.py

@@ -163,7 +163,7 @@ class Tape(object):
     __slots__ = ["_blocks", "_tf_tensors", "_tf_added_blocks", "_nodes",
                  "_tf_registered_blocks", "_bar", "_package_data",
                  "_checkpoint_manager", "latest_checkpoint",
-                 "_eagerly_checkpoint_outputs"]
+                 "_eagerly_checkpoint_outputs", "_recompute_count"]
 
     def __init__(self, blocks=None, package_data=None):
         # Initialize the list of blocks on the tape.
@@ -182,6 +182,8 @@ def __init__(self, blocks=None, package_data=None):
         self._checkpoint_manager = None
         # Whether to store the adjoint dependencies.
         self._eagerly_checkpoint_outputs = False
+        # A counter for the number of tape recomputations.
+        self._recompute_count = 0
 
     def clear_tape(self):
         """Clear the tape."""
@@ -196,6 +198,11 @@ def latest_timestep(self):
         """The current time step to which blocks will be added."""
         return max(len(self._blocks.steps) - 1, 0)
 
+    @property
+    def recompute_count(self):
+        """The number of times the tape has been recomputed."""
+        return self._recompute_count
+
     def end_timestep(self):
         """Mark the end of a timestep when taping the forward model."""
         if self._checkpoint_manager:

tests/firedrake_adjoint/test_solving.py

@@ -28,7 +28,7 @@ def J(f):
     J0 = J(f)
     rf = ReducedFunctional(J0, Control(f))
     assert_approx_equal(rf(f), J0)
-
+    assert rf.tape.recompute_count == 1
     _test_adjoint(J, f)
 
 
@@ -298,6 +298,8 @@ def test_two_nonlinear_solves():
     J = assemble(dot(u1, u1)*dx)
     rf = ReducedFunctional(J, c)
     assert taylor_test(rf, ui, Constant(0.1)) > 1.95
+    # Taylor test recomputes the functional 5 times.
+    assert rf.tape.recompute_count == 5
 
 
 def convergence_rates(E_values, eps_values):