>>>>> "Eric" == Eric Fosler-Lussier <fosler@...> writes:

Eric> On Oct 24, 2007, at 8:53 PM, David Johnson wrote:

>>
Eric> Or, does someone have some code to munge the weight files
Eric> for me?
>> I think it's a weight-file-munging exercise. Note that the new
>> weightfile format is matlab v4-compatible, so the munging
>> should be pretty straightforward. qncopywts will do this one
>> day!!

Eric> aha. I still have old rap3 weights (from qnstrn) so I can
Eric> use qncopywts to convert to matlab. I guess I could hack
Eric> the rap3 weights file rather easily (it's just lopping off
Eric> the end, if I remember) -- will qnmultifwd accept hacked
Eric> rap3 weights for a 2-layer network? Or is rap3 meant to
Eric> really only represent 3-layer networks?

I decided to let rap3 weight files die slowly in their 3-layered state
and not extend them.

David.

Eric> I have a wacky idea that I want to try out but it
Eric> presupposes getting the output from the hidden layer. I can
Eric> write the code to do this, but I figured that someone must
Eric> have done this already (particularly, maybe Barry as part of
Eric> the HATS experiments?).
>>
Eric> Also on my wish list: has anyone interfaced the feacat HTK
Eric> reading code with quicknet? Looks like it would be an easy
Eric> job to make a "htklist" input given the presence of
Eric> "srilist" but I don't want to duplicate effort.
>> I'm not aware of an htklist type. I had in the back of my mind
>> that there were issues with "srilist" - maybe something to do
>> with writing srilist files not working in feacat, or maybe
>> something more relevant?? Indeed I can't remember doing
>> anything with srilist files and qn* myself.
Eric> What I had meant was that you can convert via feacat between
Eric> pfiles and lists of htk files (or vice versa) using

Eric> feacat -ip pfile -op htk -olist output.list input.pf

Eric> feacat -ip htk -op pfile -o output.pf -lists input.list

Eric> (it isn't "htklist" per se as an argument, but using htk as
Eric> an output type implies lists of files). I saw that srilist
Eric> is an option to qnmulti{trn,fwd} so I was hoping that this
Eric> worked as advertised (i.e. not having to convert to pfiles
Eric> to train). I think having the ability to train directly
Eric> from htk files would be extremely useful for compatibility.

Eric> I'd be willing to take a crack at this but it's probably not
Eric> going to happen before January at the rate I'm going.

Eric> [BTW -- we have plans to port our Conditional Random Field
Eric> recognition software over to C++ using the qn libraries.
Eric> The current java version understands pfiles, but it would be
Eric> nice to have more interoperability. Again, not going to
Eric> happen soon.]
>> Be aware that there's a big in v3_20 (and earlier) that causes
>> problems if there are a lot of reject frames. Specifically if
>> the fraction of frames rejected ever passes the fraction of
>> work being done by a given thread (e.g. 12.5% reject with 8
>> threads). There's been a fix around for a while but I haven't
>> had time to release it (or merge Arlo's recent improvements).

Eric> Noted. Thank goodness I'm an accepting kind of guy, and
Eric> don't tend to reject much. :-)

Eric> (Actually, come to think of it, we may have seen this
Eric> bug.... hmmm....)

Eric> Thanks for the feedback, -Eric





Eric> Yahoo! Groups Links