QMR style net crashes

The engine.
Post Reply
Jochen0x90h
Posts: 19
Joined: Thu Jul 11, 2013 1:20 pm

QMR style net crashes

Post by Jochen0x90h »

Hi!

I have a QMR style test network. It can be loaded, but when evaluating it it crashes. I tested different algorithms. Do you want to have a look at it?
For some reason I can't attach the file, it always says that it is empty. The zipped version is 33kb. Can I send it by mail?

Jochen
Attachments
qmrtest.zip
(33.53 KiB) Downloaded 319 times
Last edited by Jochen0x90h on Fri Aug 23, 2013 5:23 pm, edited 1 time in total.
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: QMR style net crashes

Post by shooltz[BayesFusion] »

Jochen0x90h wrote:I have a QMR style test network. It can be loaded, but when evaluating it it crashes. I tested different algorithms. Do you want to have a look at it?
For some reason I can't attach the file, it always says that it is empty. The zipped version is 33kb. Can I send it by mail?
We had a temporary disk problem - try attaching the file again.
Jochen0x90h
Posts: 19
Joined: Thu Jul 11, 2013 1:20 pm

Re: QMR style net crashes

Post by Jochen0x90h »

Maybe some nodes have too many parents. If I restrict parent count (e.g. to 3 as a first test) it works.
for parent count restricted to 20 it starts to fail. Interesting is that Likelihood Sampling takes very long but terminates, but EPIS-Sampling crashes after some time. I wonder why this is so because sampling based algorithms should always succeed but only the quality of the result can degrade. I attached the network.
Attachments
qmr20.zip
(30.81 KiB) Downloaded 305 times
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: QMR style net crashes

Post by shooltz[BayesFusion] »

The key issue here is the following: the noisy features in SMILE are mostly for ease of modelling and not for the efficiency of inference. Almost all parts of the code, including EPIS and exact inference, operate on the CPTs obtained from noisy tables.

qmrtest.xdsl has a node with 241 parents. The total number of CPT elements is 1772567474; each of these is 8-byte double precision number. This requires over 13 gigabytes of storage just of the model itself. The additional data structures (depending on the inference algorithm) will need much more.

qmr20.xdsl will take 0.6 gigabytes for the CPTs. On my machine the EPIS code crashes during the initialization due to its additional memory requirements. I was able to run the exact inference, though.

If you are interested only in the beliefs of the subset of your node, try marking them as targets. This may lead to simpler decomposition into network fragments by the relevance layer and reduce the memory pressure.
Jochen0x90h
Posts: 19
Joined: Thu Jul 11, 2013 1:20 pm

Re: QMR style net crashes

Post by Jochen0x90h »

I already thought of this possibility that you calculate with the "unrolled" CPTs. So at first I can continue with my own loopy belief propagation and then see if I can verify the result by setting the nodes with highest probability as targets in SMILE. Maybe loopy belief propagation is already sufficient, but maybe I have to enhance it with EPIS-BN. Do you think it is possible to implement EPIS-BN based on the paper of your institute? Or do you even plan to optimize the noisy nodes so that they don't need the full cpt? We could even swap code: I give you the code of loopy belief propagation that works on noisy max parameters (without full CPT) and you give me the source of EPIS-BN 8)
Post Reply