Moved the method that retracts all variables with a given delta to Ob…

…jective (#214)

* Moved the method that retracts all variables with a given delta to Objective.

* Renamed step_optim_vars as retract_optim_vars.

theseus/core/objective.py

@@ -503,3 +503,20 @@ def to(self, *args, **kwargs):
         self.dtype = dtype or self.dtype
         if self._vectorization_to is not None:
             self._vectorization_to(*args, **kwargs)
+
+    def retract_optim_vars(
+        self,
+        delta: torch.Tensor,
+        ordering: Iterable[Manifold],
+        ignore_mask: Optional[torch.Tensor] = None,
+        force_update: bool = False,
+    ):
+        var_idx = 0
+        for var in ordering:
+            new_var = var.retract(delta[:, var_idx : var_idx + var.dof()])
+            if ignore_mask is None or force_update:
+                var.update(new_var.data)
+            else:
+                var.update(new_var.data, batch_ignore_mask=ignore_mask)
+            var_idx += var.dof()
+        self.update_vectorization()

theseus/optimizer/linear/linear_optimizer.py

@@ -69,17 +69,10 @@ def _optimize_impl(
                 warnings.warn(msg, RuntimeWarning)
                 info.status[:] = LinearOptimizerStatus.FAIL
                 return info
-        self.retract_and_update_variables(delta)
+        self.objective.retract_optim_vars(
+            delta, self.linear_solver.linearization.ordering
+        )
         info.status[:] = LinearOptimizerStatus.CONVERGED
         for var in self.linear_solver.linearization.ordering:
             info.best_solution[var.name] = var.data.clone().cpu()
         return info
-
-    # retracts all variables in the given order and updates their values
-    # with the result
-    def retract_and_update_variables(self, delta: torch.Tensor):
-        var_idx = 0
-        for var in self.linear_solver.linearization.ordering:
-            new_var = var.retract(delta[:, var_idx : var_idx + var.dof()])
-            var.update(new_var.data)
-            var_idx += var.dof()

theseus/optimizer/nonlinear/nonlinear_optimizer.py

@@ -262,8 +262,11 @@ def _optimize_loop(
                 step_size = self.params.step_size
                 force_update = False
 
-            self.retract_and_update_variables(
-                delta, converged_indices, step_size, force_update=force_update
+            self.objective.retract_optim_vars(
+                delta * step_size,
+                self.linear_solver.linearization.ordering,
+                ignore_mask=converged_indices,
+                force_update=force_update,
             )
 
             # check for convergence
@@ -376,22 +379,3 @@ def _optimize_impl(
     @abc.abstractmethod
     def compute_delta(self, **kwargs) -> torch.Tensor:
         pass
-
-    # retracts all variables in the given order and updates their values
-    # with the result
-    def retract_and_update_variables(
-        self,
-        delta: torch.Tensor,
-        converged_indices: torch.Tensor,
-        step_size: float,
-        force_update: bool = False,
-    ):
-        var_idx = 0
-        delta = step_size * delta
-        for var in self.linear_solver.linearization.ordering:
-            new_var = var.retract(delta[:, var_idx : var_idx + var.dof()])
-            if force_update:
-                var.update(new_var.data)
-            else:
-                var.update(new_var.data, batch_ignore_mask=converged_indices)
-            var_idx += var.dof()

theseus/theseus_layer.py

@@ -218,8 +218,8 @@ def backward(ctx, *grad_outputs):
         with torch.no_grad():
             bwd_optimizer.linear_solver.linearization.linearize()
             delta = bwd_optimizer.linear_solver.solve()
-            bwd_optimizer.retract_and_update_variables(
-                delta, None, 1.0, force_update=True
+            bwd_optimizer.objective.retract_optim_vars(
+                delta, bwd_optimizer.linear_solver.linearization.ordering
             )
 
         # Compute gradients.