I might be the only person ignorant enough to make this mistake, but I accidentally initialized the walkers all at the same starting location, and it took a really long time for me to figure out this was the issue.
I was trying to get it working in parallel and the program got hung up before even displaying that the Sampling progress was stuck at 0%, so I thought I was having some problem with MPI. I totally missed the warning in the FAQ that if the first few iterations take too long, it could be because the walkers started too close together.
My request is to add a check in the EnsembleSampler that the starting points are not equal, kind of like the check that all the starting logP's are finite:
|
if not np.all(np.isfinite(Z)): |
|
raise ValueError('Invalid walker initial positions! \n' + |
|
'Initialise walkers from positions of finite log probability.') |
Plus maybe an extra warning like "Do not initialize the walkers too close together!" in the FAQ's "How should I initialize the positions of the walkers?" section, but obviously, there's only so much you can do to prevent user error.
I might be the only person ignorant enough to make this mistake, but I accidentally initialized the walkers all at the same starting location, and it took a really long time for me to figure out this was the issue.
I was trying to get it working in parallel and the program got hung up before even displaying that the Sampling progress was stuck at 0%, so I thought I was having some problem with MPI. I totally missed the warning in the FAQ that if the first few iterations take too long, it could be because the walkers started too close together.
My request is to add a check in the EnsembleSampler that the starting points are not equal, kind of like the check that all the starting logP's are finite:
zeus/zeus/ensemble.py
Lines 494 to 496 in 1abdf08
Plus maybe an extra warning like "Do not initialize the walkers too close together!" in the FAQ's "How should I initialize the positions of the walkers?" section, but obviously, there's only so much you can do to prevent user error.