I've found a couple of bugs in SNNS relating to random-seed behaviour.. Both of these are upstream issues and I've sent a bug report to them. 1. When using batchman (assume also interactive tools), explicitly setting the random seed has no effect. 2. If using JE with no random seed set, an error will occur on initialization. I'm attaching a patch which corrects both issues. Reproducible: Always Steps to Reproduce:
Created attachment 155537 [details] patch proposed to upstream
phosphan is also upstream! Patrick, might want to take a look?
Thanks for the patch! It will take some time to generate new patches on berliOS, please be patient.
Looks like we can integrate this one step further upstream. Stay tuned.
There are considerations, the SNNS 4.3 to publish? Are these errors and the errors of http://developer.berlios.de/projects/snns-dev/ already integrated (in 4.3)? Since I now something new in the SNNS must implement, otherwise I would be there to watch.
My old Page: http://www1.inf.tu-dresden.de/~tr13/
(In reply to comment #5) > There are considerations, the SNNS 4.3 to publish? Are these errors and the > errors of http://developer.berlios.de/projects/snns-dev/ already integrated (in > 4.3)? Sorry for the late reply. The snns-dev patches should already be integrated in 4.3. I hope I find the time to make an ebuild soon.
The official SNNS 4.3 tarball includes some binaries by mistake, making it twice as big as necessary. Upstream (not me) did not react to the hint so far. Since it does not have any advantage I am aware of over a well-patched 4.2, I think it is useless to do a version bump right now.
(In reply to comment #8) > Since it does not have any advantage I am aware of over a well-patched 4.2, I > think it is useless to do a version bump right now. Is there a real difference in substance between the verision?
(In reply to comment #9) > (In reply to comment #8) > > Since it does not have any advantage I am aware of over a well-patched 4.2, I > > think it is useless to do a version bump right now. > > Is there a real difference in substance between the verision? It's been a while (almost a year) since I checked and can't remember exactly if there was something interesting in 3) of this list: http://www.ra.cs.uni-tuebingen.de/SNNS/newfeatures.html No idea what the "installation configuration files" were. But reading through this bug again, I think it's maybe time to get Ian's patch integrated into the snns-dev stuff as long as upstream is not reacting. Don't hold your breath.
checking up on this package. Is it ready to be bumped?
(In reply to comment #11) The upstream package hasn't changed since comment #9 was written, so as it was pointed out there I still don't think there's much point in doing a version bump -- gentoo's snns-4.2 already has all of the patches that 4.3 includes, I believe, so there really isn't a difference.
Nothing new, everything calm...
Uh.. given that the original post was about a random-seed fix, which is very much valid for snns-4.2, could that get integrated please instead of being set to WONTFIX? :) Or has it been integrated already into snns-4.2-r8 and I just missed it somewhere?
... yeah, i'd like to reopen this.. The part of this patch that keeps a memory of random weights was not applied in -r8. INIT_randomizeWeights explicitly calls krui_setSeedNo(0) on first run, which by the old definition will call srand48 with a timestamp, wiping out the previously set seed. As such I cannot control the randomized weights via a seed setting, and therefore cannot reproduce a training run. If the store-the-seed-in-memory method is undesirable, perhaps keeping around something that will indicate whether the seed has been set (and not call it again) would be better??
re-filing as per the original and as yet unfixed bug.
(In reply to comment #14) > Uh.. given that the original post was about a random-seed fix, which is very > much valid for snns-4.2, could that get integrated please instead of being set > to WONTFIX? :) > > Or has it been integrated already into snns-4.2-r8 and I just missed it > somewhere? Sorry... made this mistake because I confounded it with another random seed ralated issue and the whole sad 4.3 story. Will have a closer look around next weekend.
Included the patch in -r9. Sorry that it took me so long.