Optimizer class tracks a cnt counter to decide when to apply the minibatch update. However, it appears that the counter is incremented not only in the _learn driver loop (
), but also in the concrete optimizer class
update_general methods (
), which is called once for each layer for each step.
Is this intended behavior? The minibatch update can trigger in some middle layer and reset the counter, rather than every mini-batch-size number of observables backprops.
Optimizer class tracks a
cntcounter to decide when to apply the minibatch update. However, it appears that the counter is incremented not only in the_learndriver loop (vectorflow/src/vectorflow/optimizers.d
Line 138 in d26cd9d
update_generalmethods (vectorflow/src/vectorflow/optimizers.d
Line 307 in d26cd9d
Is this intended behavior? The minibatch update can trigger in some middle layer and reset the counter, rather than every mini-batch-size number of observables backprops.