I tried to use Pytorch(v=0.4.1) together with Joblib in order to parallelize the sampling process in a RL task. Setting the seed for Pytorch, Numpy, and python wasn’t enough to get reproducible experiments. In other words, Pytorch yielded me different predictions for different runs of the same algorithm in the same task with the same seeds. I am wondering if there is any known issue between Pytorch and Joblib. With Multiprocessing the results are just fine.
gcnn (gcnn) #1