val loop :
start_index:int ->
end_index:int ->
save_vars_from:Node.p list ->
checkpoint_base:string ->
?checkpoint_every:[ `iters of int | `seconds of float ] ->
(index:int -> unit) ->
unitloop ~start_index ~end_index ~save_vars ~checkpoint_base f starts a loop ranging from start_index to end_index. On each iteration f is called on the current index.
This loop is checkpointed: regularly the state of all the variables reachable from save_vars_from are saved on disk to file checkpoint_baseXXX. If such files already exist when starting the loop, the last one of this file is used to restore the variables content.
Note that this assumes that the graph is generated in the same way on both side as variables need to have the same ids in current graph and in the loaded checkpoint.