Difference between the 2 codes below (REINFORCE)

autograd
21

Screenshot%20from%202018-07-16%2016-44-26
Screenshot%20from%202018-07-16%2016-44-261920×1080 249 KB

I used Adam with lr = 1e-3
First one is CPU

22

Seems to work. Maybe you need to train a bit longer if you use non-deterministic behavior but in general it should converge too.

1 Like
23

My concern is mostly with the values, since the seed is same. I thought, the values will also be same. Does it have to do with data types?