Add autograd create graph and strategy options for AutoDiffCostFuncti…

…on. (facebookresearch#532)

theseus/core/cost_function.py

@@ -209,8 +209,10 @@ def __init__(
         cost_weight: Optional[CostWeight] = None,
         aux_vars: Optional[Sequence[Variable]] = None,
         name: Optional[str] = None,
+        autograd_create_graph: bool = True,
         autograd_strict: bool = False,
         autograd_vectorize: bool = False,
+        autograd_strategy: str = "reverse-mode",
         autograd_mode: Union[str, AutogradMode] = AutogradMode.VMAP,
     ):
         if cost_weight is None:
@@ -228,8 +230,10 @@ def __init__(
 
         self._err_fn = err_fn
         self._dim = dim
+        self._autograd_create_graph = autograd_create_graph
         self._autograd_strict = autograd_strict
         self._autograd_vectorize = autograd_vectorize
+        self._autograd_strategy = autograd_strategy
 
         # The following are auxiliary Variable objects to hold tensor data
         # during jacobian computation without modifying the original Variable objects
@@ -285,9 +289,10 @@ def _compute_autograd_jacobian(
         return autogradF.jacobian(
             jac_fn,
             optim_tensors,
-            create_graph=True,
+            create_graph=self._autograd_create_graph,
             strict=self._autograd_strict,
             vectorize=self._autograd_vectorize,
+            strategy=self._autograd_strategy,
         )
 
     def _make_jac_fn_vmap(