... Ah no problem at all. Glad you're feeling better, no need for excuses. Sorry, I was wrong not to mention that I also added a '-E <epsilon>' runtime...
231
John Langford
john_c_langford
Jul 9, 2011 3:21 pm
... The only reports I've seen of VW in use on non-x86 are for ARM systems. We can setup a debugging session, although scheduling it will probably have to wait...
230
John Langford
john_c_langford
Jul 9, 2011 3:16 pm
Oops missed this(*). It's incorporated now. I set the precision to 0.001 in the test file. -John (*) Excuse = was traveling, and then got sick while...
229
Bharath Krishnan
cqbkrishnan
Jul 9, 2011 2:37 pm
Thanks, I will test out the patch. -bharath...
228
warezio
Jul 9, 2011 5:05 am
Ok. First, I was able to reproduce. Second, it looks like problem is isolated to total_lambda, which is a double, unlike most everything else in the file which...
227
Yaroslav Halchenko
debian@...
Jul 9, 2011 1:22 am
which platform is of particular interest? I guess I could setup an interactive session so we could together visit one of the Debian machines for that platform...
226
warezio
Jul 9, 2011 1:03 am
Ok I'll check it out this weekend and try to reproduce. One thing that immediately jumps out: the large-ish value of --lda_D in both your original case and in...
225
John Langford
john_c_langford
Jul 9, 2011 12:20 am
The differences on some of these platforms are so significant that I don't believe they are just due to the use of -ffast-math. Could the random initialization...
224
John Langford
john_c_langford
Jul 9, 2011 12:16 am
I added the '-f' flag to Runtests under 'make test', which Ariel setup to do fuzzy match. This seems to address the issue on my machine. -John...
223
Bharath Krishnan
cqbkrishnan
Jul 8, 2011 11:43 pm
Hi Paul, I can reproduce something similar with the wiki1K data set. I exponentially concatenated the wiki1K.dat file till it had 16384000 lines. Model built...
222
warezio
Jul 8, 2011 8:11 pm
Ah that's good sleuthing! Can you share your data set? -- p...
221
warezio
Jul 8, 2011 8:11 pm
Ah that's good sleuthing! Can you share your data set? -- p...
220
Bharath Krishnan
cqbkrishnan
Jul 8, 2011 7:21 pm
I tried another run with MINEIRO_SPECIAL defined but compiled with -DVW_LDA_NO_SSE. If I understand this correctly, this turns off the SSE vectorization code....
219
Bharath Krishnan
cqbkrishnan
Jul 8, 2011 6:29 pm
Hi Paul, That does seem to be the case. I tested it without the faster math approximations and now I don't see very large numbers/Inf in the model anymore. Let...
218
warezio
Jul 8, 2011 3:58 pm
My LDA code changes are the #1 suspect. Try commenting out the "#define MINEIRO_SPECIAL" in the lda_core.cc file, recompiling, and re-running with this data....
217
Bharath Krishnan
cqbkrishnan
Jul 8, 2011 3:11 pm
Hi Folks, I am testing out the latest lda code & I am seeing some weird models show up with our data sets. The older ~5.1 vw version doesn't seem to do this. ...
216
Ariel Faigon
ariel_vw
Jul 6, 2011 8:48 pm
... My case is worse than Michael's. I don't get all zeros, so the predictions look 'correct' when they are sometimes way off (especially with...
215
Ariel Faigon
ariel_vw
Jul 5, 2011 7:08 pm
Hi, Every time I upgrade boost, which seems to happen pretty frequently on Ubuntu, I need to rebuild my multiple versions of vw or they fail to start. boost...
214
joal
jyodoshi_mik@...
Jul 5, 2011 3:51 pm
Hi John, Interesting. Maybe the bit size isn't being read: cat trn |vw -b 29 -f mdl this predicts all 0's cat tst |vw -t -i mdl while this predicts something...
213
Ariel Faigon
ariel@...
Jul 5, 2011 2:39 pm
Michael, Yes, this has bit me (hard) too in the past. The -b setting changes the computation of the hash values so it must be consistent between training and...
212
John Langford
john_c_langford
Jul 5, 2011 2:33 pm
The number of bits actually is a part of the model. You can check this by downloading the latest and using --readable_model <foo> But, perhaps there is some...
211
John Langford
john_c_langford
Jul 5, 2011 2:02 pm
... No. Making the daemon persistent is both desirable and not yet done. -John...
210
joal
jyodoshi_mik@...
Jul 4, 2011 8:30 pm
Hi, I've been using vw lately -nice software. I spent some hours figuring out that setting a custom feature bit size in training requires it to be explicitly...
209
Matt Hoffman
flashrefuse
Jul 4, 2011 3:25 pm
Hi Shaohua, When using a relatively large number of topics, most words in most topics should be close to the prior smoothing level of 0.1. But not all—some...
208
jeffrilee
Jul 4, 2011 7:40 am
EC2 reserves around 9G space for each user. I guess it's not caused by disk full, cuz after vw terminates, I still get 4G free spaces. But I found EC2 is so...
207
jeffrilee
Jul 4, 2011 7:36 am
Thank you Ariel! I've installed linux on my PC, run vw and got similar results. What puzzled me is most words in most topics get weights of 0.1. Only a few...
206
Matt Hoffman
flashrefuse
Jul 3, 2011 2:57 pm
... Maybe this is the explanation? How much storage does EC2 let you use? Matt...
205
Ariel Faigon
ariel_vw
Jul 3, 2011 9:50 am
FWIW: I can't reproduce the crash. I downloaded the training set and ran it using the latest code (current master from...
204
Matt Hoffman
flashrefuse
Jul 2, 2011 3:13 pm
Hi Shaohua, From your email it's not easy to answer this question. I recommend trying to run this on your own computer and seeing if it works there (keeping...
203
jeffrilee
Jul 2, 2011 10:44 am
Dear all, I used vw on this dataset (created by me): http://www.ntu.edu.sg/home2009/lish0018/venues-num.zip Each line represents an author in an academic...
