This is fix of IterativeLoss. (hot fix #114 is closed.)
Issue: a bug of IterativeLoss
x_dict.update({value: x_dict[key]}) at line:156 register key-variable's value as value-variable's value. And key-variable remains.
At next step, key-variable are sometimes overwritten but sometimes not. For example, AddLoss copies an input dictionary, and overwrite old one to updated one.
the key-variable is not updated. So the effect of reccurence is ignored.
Solution
delete future variable by replace_dict_keys at the step of IterativeLoss.
get only updated variables from distribution.sample or loss.eval and override them to dictionary of all variables.
This is fix of IterativeLoss. (hot fix #114 is closed.)
Issue: a bug of IterativeLoss
x_dict.update({value: x_dict[key]})
at line:156 registerkey
-variable's value asvalue
-variable's value. Andkey
-variable remains.key
-variable are sometimes overwritten but sometimes not. For example,AddLoss
copies an input dictionary, and overwrite old one to updated one.key
-variable is not updated. So the effect of reccurence is ignored.Solution