It would be nice to support the numpy.random methods get_state() and set_state() for the distributed system. This would let the user save and restore the global prng generator state, even when they did not use a deterministic seed to begin with.
For the distributed system, this would presumably work with a list of generator states, one per engine, since these are in general different. I started to try an implement this in PR #253, but the data being communicated is fairly large (about 624 integers per engine), which was slow. This specific usage was considered less important than what #253 provides, so this issue is created to note the desire for later.
kwmsmith
commented
10 years ago
It would be nice to support the
numpy.randommethodsget_state()andset_state()for the distributed system. This would let the user save and restore the global prng generator state, even when they did not use a deterministic seed to begin with.For the distributed system, this would presumably work with a list of generator states, one per engine, since these are in general different. I started to try an implement this in PR #253, but the data being communicated is fairly large (about 624 integers per engine), which was slow. This specific usage was considered less important than what #253 provides, so this issue is created to note the desire for later.