A big downside in the design of Storchastic is how storch.backward uses the global state of the costs, instead of calling .backward on a cost.
Would we be able to design it like that? It would require knowing when to free the backwards buffers, and it would not use a single backwards pass. Might be inefficient...
A big downside in the design of Storchastic is how storch.backward uses the global state of the costs, instead of calling .backward on a cost.
Would we be able to design it like that? It would require knowing when to free the backwards buffers, and it would not use a single backwards pass. Might be inefficient...