I am trying to add custom gradient support to the Java bindings (tensorflow/java#292). This is currently being prevented by the fact that TF_AddGradientsWithPrefix requires a lock on the graph, but so do any operation creation methods (i.e. TF_NewOperation), so it is impossible to create ops in gradient functions.
Is there a way I can get around this? I am considering calling TF_NewOperationLocked directly but I assume the locks are there for a reason.
I am trying to add custom gradient support to the Java bindings (tensorflow/java#292). This is currently being prevented by the fact that
TF_AddGradientsWithPrefixrequires a lock on the graph, but so do any operation creation methods (i.e.TF_NewOperation), so it is impossible to create ops in gradient functions.Is there a way I can get around this? I am considering calling
TF_NewOperationLockeddirectly but I assume the locks are there for a reason.