@sandeep-iitr there have been few issues about the meaning of n_iterations in parallel mode. Intuitively (since all other tuning libraries have this behavior) n_iterations should be same as number of objective function evaluation but Mango does n_iterations x batch_size.
I propose that we deprecate n_iterations and add a new parameter: n_samples Number of function evaluation would always be == n_samples in serial and parallel cases. In parallel case the number of batches is calculated as n_samples / batch_size.
We would add a deprecation warning to n_iterations parameter to warn users about the change and update all docs and examples to use n_samples. This way it won't break existing usage.
Hi Mohit, Thanks for raising.
I agree with your observations. More recently, we have seen many users asking for clarifications on the total number of samples vs the number of iterations parameter.
@sandeep-iitr there have been few issues about the meaning of
n_iterations
in parallel mode. Intuitively (since all other tuning libraries have this behavior)n_iterations
should be same as number of objective function evaluation but Mango doesn_iterations x batch_size
.I propose that we deprecate
n_iterations
and add a new parameter:n_samples
Number of function evaluation would always be== n_samples
in serial and parallel cases. In parallel case the number of batches is calculated asn_samples / batch_size
.We would add a deprecation warning to
n_iterations
parameter to warn users about the change and update all docs and examples to usen_samples
. This way it won't break existing usage.I can make all the changes needed.