In the last post I asked about a "normal" training accuracy, and thanks Arlo that I'm in the right way. So I started to use QuickNet for our own job, but only...
Hi Xuchen, (CC'd to the list, for reference) I looked at your log file and it seems there's no bugs. Moreover, your results seem perfectly "normal" ... when...
Arlo Faria
arlo@...
May 15, 2007 4:40 am
87
Thank you Arlo. In order to decide which is the best hidden unit size, I read the paper "Size Matters" before. The whole job contains 40M frames, with 39 input...
Your input layer has 39*9=351 units, so the formula would suggest about 4000 hidden units. If this is too big/slow, I would recommend training on a subset of...
Arlo Faria
arlo@...
May 15, 2007 5:45 am
89
Hi, I trained a net for 4000 utterances and then forward all 4000 utterances using qnsfwd. But I found that the first 4000 utterances contain more frames than...
It comes from the fact that you have a context window of 9 frames (since you don't have the qnstrn script here, it's hard to say which one, but if the...
Eric Fosler-Lussier
fosler@...
May 22, 2007 3:07 pm
91
Thank you, Eric. When using quicknet, I didn't pad either feature file or ilab pfile. So I always set hardtarget_window_offset=0 when using qnstrn and qnsfwd....
... Yes, and train with hardtarget_window_offset=4, I would recommend. This will skip the first 4 labels (which is what you want). Otherwise, with...
Eric Fosler-Lussier
fosler@...
May 23, 2007 1:31 am
93
Well, I didn't notice the -deslen option of feacat/labcat before. Maybe there's no example in the manpage so I never thought to try it... Anyway, I learned a...
It's actually a weird feature that I had Dan Ellis put in for me. I don't know if anyone other than me uses it. :-) -EFL...
Eric Fosler-Lussier
fosler@...
May 23, 2007 2:08 am
95
I use the -deslen option all the time... maybe too much! I have to coordinate feature and label streams from ICSI, SRI, HTK, and ESPS tools, and they never...
Arlo Faria
arlo@...
May 23, 2007 8:51 am
96
With the help of many kind people (thanks a lot!) in this mail list, I finally got the correct posteriors with correct and matched frame numbers. So I guess...
Congratulations on getting a network trained -- that's the hardest part! I'm assuming that you're trying to generate Tandem-style features, rather than...
Arlo Faria
arlo@...
May 28, 2007 1:30 am
98
Thank you very much, Arlo. Yes, I do want to generate tandem features. I followed your step and wrote an awk script to compute how many components contain 95%...
By the way, if I train a network using "softmax" as the output type, can I specify the output type of qnsfwd to "linear" to get the linear activations? I mean,...
Yes, that's correct: you can set softmax during training and linear during the forward pass. You could also train with linear outputs, in which case I believe...
Arlo Faria
arlo@...
May 29, 2007 3:04 am
101
The long-awaited version v3_20 of the QuickNet MLP training program and libraries has been released and installed in /u/drspeech/ at ICSI. Apologies for the...
Version 1.01 of the feacat package has been released, tagged in the CVS archive and installed in /u/drspeech/ at ICSI. feacat includes several versatile...
SHORT: Where one can obtain dpwelib sources? LONGER: - feacat 1.01 requires dpwelib and QuickNet3 libs to compile - QN version 3.20 compiled OK (thanks!),...
... Petr> SHORT: Where one can obtain dpwelib sources? LONGER: - Petr> feacat 1.01 requires dpwelib and QuickNet3 libs to compile Petr> - QN version 3.20...
The configure script error you mentioned is probably caused by Tcl/Tk, if your linux box is quite new. Use the following script before configure: cp...
I use feacalc-0.90. And feacalc stops whenever the pfile is >= 2GB. I remembered that there maybe some historical issues that some tools from ICSI don't...
I use pfile_norm with option "-slf LENGTH_FILE" when do per speaker normalization. Whenever LENGTH_FILE is a little large, I got the following error: *** glibc...
... yaoxuchen> I use pfile_norm with option "-slf LENGTH_FILE" when do yaoxuchen> per speaker normalization. yaoxuchen> Whenever LENGTH_FILE is a little large,...
... yaoxuchen> I use feacalc-0.90. And feacalc stops whenever the yaoxuchen> pfile is >= 2GB. I remembered that there maybe some yaoxuchen> historical issues...
Xuchen, your configure trick succeded; you were right also with gcc version. Meanwhile I found a patch for dpwelib-2002-10-23 (again from Olivier Bornet) and...
Having problems with message search?
Fill out this form to ensure your group is one of the first to be migrated to the new message search system.
