H20 Deep learning... bug when set to reproducible and random seed?

hughesfleming68

I am not sure what is happening but I am using the H20 Deep Learner. It works fine until I set it to reproducible. Generally this operator is not the fastest but when set to use one thread and random seed, it starts spitting out results at high speed as if it is not really doing anything except generating random predictions.  I don't remember it working this way in the past. Perhaps some can confirm that everything is normal.

Edit- This is within a sliding window validation and I am logging performance so I can see how fast it is generating results.

Thanks,

Alex

sgenzer

Hi Alex - can you post your XML and data so we can see? Your rapidminer-studio.log file would also be useful - you can send via PM if you want.

Scott

hughesfleming68

Thanks Scott, I will zip up a processes and send it to you in the morning to see if you can duplicate it. I have a couple of other issues like the select attributes operator not seeing all my attributes or having difficulty with metadata. I will fire up a VM and reinstall and see if my problems go away.

regards,

Alex

sgenzer

perfect. Thx Alex.

hughesfleming68

I started up RM82 on Windows Server 2016 using the full Windows installation rather than the Linux install to avoid any problems that an alternative jvm might cause. The problem is the same. Setting the process to reproducible (single thread) results in a massive speed up. My guess is that the results are random. A quick test is to start the process with the switch on and off and then review the log in the results. GBT does not do this so this seems to be limited to the Deep Learner in H20.

regards,

Alex