I like this new feature as simplify the code a lot, however I have few doubts for which I could not find much information around.
<offset= ... ,multiplier= ... ><lower= .., upper = .. >seems to not be accepted. Is there anything I am missing?
It would be great to be able to specify a vector for offset
vector<offset = vector[N] >[N] alpha
Is there any plan to integrate this?
The reason being that if I have a big model for third party execution of which 10K+ parameters (encoded as vector) have a value of roughly a known value I could avoid to waste the initial 50 iterations in order to find the scale
Would the second best thing be to set initial values?
If I set
offsetto speed up adaptation, I imagine I don’t have to specify initial values, or this still plays a role?
For variational inference is there anything to worry about in the use of <offset= … ,multiplier= … >?