Open vitchyr opened 5 years ago
Hi, I have the same problem, have you sovled this?
I run into this error as well.
Same here, but for the HandManipulatePen env, though it prints out the first log table just fine.
I've temporarily fixed the issue in my custom environment by having the success rate over 100 rather than 1 (ie. multiplying the success rate by 100 in rollout.py). The issue lies somewhere in mpi_moments.py when calculating the meansqdiff with non-zero success rate with multiple cpus. Hope this is useful.
It seems like an issue related to x.dtype
. Probably workers are for some reason generated a different x.dtype
and MPI is not able to reconcile them. A workaround is to force the dtype of localsum in line 12 to be np.float64
. The right fix would be figure out why different dtypes.
The above work around works for me thanks
time mpirun -np 8 meams num-cpu=8? and if i want use mpi,the command "mpirun "must be added ?
time mpirun -np 8 meams num-cpu=8? and if i want use mpi,the command "mpirun "must be added ?
Yes, that's how it works. Check out the mpi documentation about mpirun, but you need to add it to use mpi.
The following command runs fine:
However, if I try changing the environment to the FetchPush-v1 or FetchPickAndPlace-v1, I get the following error: When trying to run multiple MPI threads
These different environments work for me if I run them without MPI.
I am using anaconda. My Python version is 3.6.2 and this is the output of
pip freeze
: