Workers dynamically join & leave about parameter_server

[mli]

Hi Lisen,

Please have a check at the dev branch, that feature is partially supported.
But I need more tests.

Best
Mu
On Sun, Mar 29, 2015 at 11:25 PM Lisen Mu [email protected] wrote:

Hi,

In data parallelism, it is reasonable(or is it? ) to allow workers start
to train a model replica at anytime, for 2 reasons: a) worker's resource
may not all available at the start of training, b) for better convergence,
1 worker could start first to train a while to give a kickstart before
other workers join the group ( mentioned here:
http://research.google.com/archive/large_deep_networks_nips2012.html ).

Currently it seems that all servers/workers should be initiated before
anyone of them start to run. And it seems to be controlled by PostOffice
and can not easily be customized with my Scheduler or Server.

Thanks,

—
Reply to this email directly or view it on GitHub
#17.

[immars]

Thanks again @mli !