Very long training Time with smaller network

The front end.
mark
Posts: 179
Joined: Tue Nov 27, 2007 4:02 pm

Post by mark »

Although the bug only occurred in one particular setting (otherwise I would have caught it earlier), I do recommend to retrain the networks.

One disadvantage of your approach to soft evidence is that you will need to perform inference multiple times, but that may not be a problem.
Christian
Posts: 44
Joined: Wed Nov 28, 2007 12:32 pm

Post by Christian »

Yes, calculating would be a disadvantage. But as I will have precalculated every possible evidence settings (for speed enhancement) it is not that problem for me.
mark
Posts: 179
Joined: Tue Nov 27, 2007 4:02 pm

Post by mark »

GeNIe and most of the SMILE versions have been updated on our website. If you need a version of SMILE that is not updated yet, please let me know.
Christian
Posts: 44
Joined: Wed Nov 28, 2007 12:32 pm

Post by Christian »

Thank you for your updates. For Batch Training I am using smile.net. This one is not updated yet.

But I will start training of the same network with the same data now and will let you know how much the improovements are in this sample. I don't know how much it can take. This network took the last time 1800 minutes.
mark
Posts: 179
Joined: Tue Nov 27, 2007 4:02 pm

Post by mark »

Smile.NET is available now.
Christian
Posts: 44
Joined: Wed Nov 28, 2007 12:32 pm

Post by Christian »

Thank you!
Christian
Posts: 44
Joined: Wed Nov 28, 2007 12:32 pm

Post by Christian »

Hello Marc,

I want to let you know, that the Genie Training is now finished.

The networks, Training data, computer usage - everything was the same.

Old Training Time: 1864 minutes
New Training time: 1348 minutes

It is an improvement of 28%. This way I will get about every third network training for free :-)

If you know a way to improve the em-algorithm even further, I would really appreciate it. Once You have a new version I will rerun the same tests.

Thank you very much!

Christian
mark
Posts: 179
Joined: Tue Nov 27, 2007 4:02 pm

Post by mark »

That's a nice improvement and we will try to make it even faster. Thanks for testing!
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Post by shooltz[BayesFusion] »

Christian wrote:Old Training Time: 1864 minutes
New Training time: 1348 minutes
If you're still interested in EM, most recent GeNIe build (2.0.3092.0) should perform even better.
Christian
Posts: 44
Joined: Wed Nov 28, 2007 12:32 pm

Post by Christian »

Hi,

thank you for your updates. Currently I am working on a different project, but as soon as I get back to the other one I will let you know the differences.

Regards,

Christian
Post Reply