Menu
▾
▴
ghmm-list — Mailing list for all things ghmm.
You can subscribe to this list here.
| 2004 |
Jan
(1) |
Feb
|
Mar
(13) |
Apr
|
May
(1) |
Jun
|
Jul
(10) |
Aug
(6) |
Sep
|
Oct
|
Nov
(2) |
Dec
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2005 |
Jan
(5) |
Feb
(6) |
Mar
(3) |
Apr
(2) |
May
(7) |
Jun
(7) |
Jul
(2) |
Aug
(8) |
Sep
(7) |
Oct
(11) |
Nov
|
Dec
(14) |
| 2006 |
Jan
(13) |
Feb
(2) |
Mar
(13) |
Apr
(21) |
May
(3) |
Jun
(5) |
Jul
(4) |
Aug
|
Sep
(3) |
Oct
(6) |
Nov
(15) |
Dec
(4) |
| 2007 |
Jan
(2) |
Feb
(1) |
Mar
(3) |
Apr
(1) |
May
(19) |
Jun
(4) |
Jul
|
Aug
(1) |
Sep
(1) |
Oct
(2) |
Nov
(7) |
Dec
(2) |
| 2008 |
Jan
(2) |
Feb
|
Mar
(4) |
Apr
(19) |
May
(13) |
Jun
(22) |
Jul
(4) |
Aug
|
Sep
(3) |
Oct
(24) |
Nov
(4) |
Dec
|
| 2009 |
Jan
(5) |
Feb
(16) |
Mar
(3) |
Apr
(7) |
May
(2) |
Jun
|
Jul
|
Aug
(4) |
Sep
|
Oct
(1) |
Nov
|
Dec
|
| 2010 |
Jan
(2) |
Feb
(4) |
Mar
|
Apr
|
May
|
Jun
|
Jul
(22) |
Aug
(2) |
Sep
(4) |
Oct
(22) |
Nov
(4) |
Dec
(1) |
| 2011 |
Jan
|
Feb
|
Mar
(5) |
Apr
(1) |
May
(17) |
Jun
|
Jul
|
Aug
(1) |
Sep
|
Oct
(1) |
Nov
(3) |
Dec
|
| 2012 |
Jan
|
Feb
(1) |
Mar
(1) |
Apr
|
May
(3) |
Jun
(2) |
Jul
|
Aug
(3) |
Sep
|
Oct
|
Nov
(3) |
Dec
|
| 2013 |
Jan
(1) |
Feb
|
Mar
(1) |
Apr
(2) |
May
(1) |
Jun
(1) |
Jul
(7) |
Aug
(5) |
Sep
|
Oct
(1) |
Nov
|
Dec
|
| 2014 |
Jan
(4) |
Feb
|
Mar
(1) |
Apr
(2) |
May
(2) |
Jun
|
Jul
(1) |
Aug
|
Sep
|
Oct
|
Nov
|
Dec
(1) |
| 2015 |
Jan
(1) |
Feb
|
Mar
(2) |
Apr
(3) |
May
(2) |
Jun
(2) |
Jul
|
Aug
|
Sep
(1) |
Oct
|
Nov
|
Dec
|
| 2020 |
Jan
|
Feb
(1) |
Mar
|
Apr
|
May
|
Jun
|
Jul
|
Aug
|
Sep
|
Oct
|
Nov
|
Dec
|
|
From: Dimitris D. <ddi...@gm...> - 2020-02-19 03:27:37
|
Hi, I'm looking to train an HMM model (1st order but if 2nd order or more is available that would be great) that has a custom emission probability distribution. Is that supported in GHMM ? If so, are there any code examples posted somewhere, or any specific functions I should look for in the API ? Regards, Dimitris |
|
From: Mustapha Km <km....@gm...> - 2015-09-26 15:48:18
|
hi everybody I was wondering how can we introduce a poisson distribution into the ghmm library and do some learning using baum_welsh with that distribution |
|
From: BingXin Lu <sou...@gm...> - 2015-06-26 02:43:04
|
Hi,
I am a newbie to ghmm.
I want to use ghmm to build a higher order two-state HMM for a DNA sequence.
But I do not know how to set the emission probabilities and transition
probabilities correctly.
For example, I want to use a 5th order HMM,
I tried to set:
num = 4 ** 6
emit = [.25] * num
But while training, it shows:
GHMM ghmm.py:148 - reestimate.c:ghmm_dmodel_baum_welch(817):
reestimate_one_step false (1.step)
Would you please tell me what is the right way to set this emission
probability vector?
Besides, I want to have two states modeled by HMM. There should be only one
transition point between the two states.
After transition from state 0 (normal) to state 1 (alien), the sequence
should stay in state 1 until the end.
I tried to set the transition probabilities as follows:
norm_alien = 1 / duration
norm_norm = 1 - 1 / duration
norm_trans = [norm_alien, norm_norm]
alien_norm = 0
alien_alien = 1
alien_trans = [alien_norm, alien_alien]
But there will be multiple transition points reported when I use 0-order
HMM.
Do you know how can my requirements be satisfied?
Thanks a lot!
Best Regards,
Bingxin
|
|
From: Stalin M. <sta...@ya...> - 2015-06-01 23:18:14
|
Hi! I'd like to use ghmm for modelling speech in a recognition task and have some problems having it to work with a basic setting: Have the following code: F = ghmm.Float() ... print pi [1.0, 0.0, 0.0, 0.0] print transition [[ 0.33333333 0.33333333 0.33333333 0. ] [ 0. 0.33333333 0.33333333 0.33333333] [ 0.33333333 0. 0.33333333 0.33333333] [ 0.5 0. 0. 0.5 ]] print gaussian_params [[[-2.3086799162957226, 0.07214405003411178, 0.23762250285550643, -0.14486415399066022, 0.26475867914248574, 0.440033594010816, -0.046810570754743014, 0.18445631270914667, 0.627266653774591, -0.1717022826765591, 0.10829636983950663, 0.06200962863341574, -0.09284077297749817], [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0],[0.03798026817242703, 1.8580771395415046, -1.5641498864922676, -0.9203836538187552, -0.5632487857647773, -0.7375090533036678, 0.2238434025115639, 0.37949919183751124, 0.40332874879336666, -0.5063901036346233, -0.11411751178201408, -0.3299903725491531, -0.11387212463813698], [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [0.30510293843156744, 1.7152793719556199, 1.014069570193344, 0.562556639529057, 0.29202438061211233, 0.5577457024268492, 0.18177792257754283, 0.37239809827293735, 0.5387617547239218, -0.00291152516540962, 0.29478802498327517, 0.17137965336187408, 0.0050087786030910035], [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [0.3333333333333333, 0.3333333333333333, 0.3333333333333333]], [[-2.3086799162957226, 0.07214405003411178, 0.23762250285550643, -0.14486415399066022, 0.26475867914248574, 0.440033594010816, -0.046810570754743014, 0.18445631270914667, 0.627266653774591, -0.1717022826765591, 0.10829636983950663, 0.06200962863341574, -0.09284077297749817], [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [0.03798026817242703, 1.8580771395415046, -1.5641498864922676, -0.9203836538187552, -0.5632487857647773, -0.7375090533036678, 0.2238434025115639, 0.37949919183751124, 0.40332874879336666, -0.5063901036346233, -0.11411751178201408, -0.3299903725491531, -0.11387212463813698], [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [0.30510293843156744, 1.7152793719556199, 1.014069570193344, 0.562556639529057, 0.29202438061211233, 0.5577457024268492, 0.18177792257754283, 0.37239809827293735, 0.5387617547239218, -0.00291152516540962, 0.29478802498327517, 0.17137965336187408, 0.0050087786030910035], [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [0.3333333333333333, 0.3333333333333333, 0.3333333333333333]], [[-2.3086799162957226, 0.07214405003411178, 0.23762250285550643, -0.14486415399066022, 0.26475867914248574, 0.440033594010816, -0.046810570754743014, 0.18445631270914667, 0.627266653774591, -0.1717022826765591, 0.10829636983950663, 0.06200962863341574, -0.09284077297749817], [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [0.03798026817242703, 1.8580771395415046, -1.5641498864922676, -0.9203836538187552, -0.5632487857647773, -0.7375090533036678, 0.2238434025115639, 0.37949919183751124, 0.40332874879336666, -0.5063901036346233, -0.11411751178201408, -0.3299903725491531, -0.11387212463813698], [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [0.30510293843156744, 1.7152793719556199, 1.014069570193344, 0.562556639529057, 0.29202438061211233, 0.5577457024268492, 0.18177792257754283, 0.37239809827293735, 0.5387617547239218, -0.00291152516540962, 0.29478802498327517, 0.17137965336187408, 0.0050087786030910035], [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [0.3333333333333333, 0.3333333333333333, 0.3333333333333333]], [[-2.3086799162957226, 0.07214405003411178, 0.23762250285550643, -0.14486415399066022, 0.26475867914248574, 0.440033594010816, -0.046810570754743014, 0.18445631270914667, 0.627266653774591, -0.1717022826765591, 0.10829636983950663, 0.06200962863341574, -0.09284077297749817], [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [0.03798026817242703, 1.8580771395415046, -1.5641498864922676, -0.9203836538187552, -0.5632487857647773, -0.7375090533036678, 0.2238434025115639, 0.37949919183751124, 0.40332874879336666, -0.5063901036346233, -0.11411751178201408, -0.3299903725491531, -0.11387212463813698], [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [0.30510293843156744, 1.7152793719556199, 1.014069570193344, 0.562556639529057, 0.29202438061211233, 0.5577457024268492, 0.18177792257754283, 0.37239809827293735, 0.5387617547239218, -0.00291152516540962, 0.29478802498327517, 0.17137965336187408, 0.0050087786030910035], [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [0.3333333333333333, 0.3333333333333333, 0.3333333333333333]]] print F <ghmm.Float object at 0x7f6c50776f90> print ghmm.MultivariateGaussianDistribution(F) <ghmm.MultivariateGaussianDistribution object at 0x7f6c50776f50> In the following step I try to create the model but it blocks the program (stays there for hours not returning -do not have any feedbak from it) model = ghmm.HMMFromMatrices(F,ghmm.MultivariateGaussianDistribution(F),transition, gaussian_params, pi) The only reference I have for using this class is in the following link: General Hidden Markov Model Library / Mailing Lists | | | | | | | | | | | General Hidden Markov Model Library / Mailing ListsOk. Then if I have time one of these days I'll send you a patch to correct that issues. | | | | View on source... | Preview by Yahoo | | | | | Is the format of the parameters correct? Best regards, Stan |
|
From: Alexander S. <sc...@cs...> - 2015-05-08 19:17:50
|
> On May 1, 2015, at 5:59 PM, Fatma Miladi <mil...@ya...> wrote: > > Hi, > In my work, i need to cluster sequencial data into groups. So, i used cygwin to compile ghmm/tools/cluster.c using this instruction : gcc cluster.c –o c1. It works well. However when i bulit it with this instruction : ./c1 test1.sqd test2.smo test3.smo, i had this message: "cluster is obsolete. If you need it rebuild the GHMM with "GHMM_OBSOLETE". Please how can i rebuild the GHMM with "GHMM_OBSOLETE? Thanks in advance. > Best regards, Check the output of ./configure —help for configuration options. As far as I remember --enable-unsupported. A, > > ---- > Fatma MILADI > PhD Student in Computer Science > Member at MIRACL laboratory > ------------------------------------------------------------------------------ > One dashboard for servers and applications across Physical-Virtual-Cloud > Widest out-of-the-box monitoring support with 50+ applications > Performance metrics, stats and reports that give you Actionable Insights > Deep dive visibility with transaction tracing using APM Insight. > http://ad.doubleclick.net/ddm/clk/290420510;117567292;y_______________________________________________ > Ghmm-list mailing list > Ghm...@li... > https://lists.sourceforge.net/lists/listinfo/ghmm-list -- See http://schliep.org/CATBox for Springer Book Dr. Alexander Schliep, Associate Professor, Tel: (848) 445-7286 sc...@cs... Lab webpage: http://bioinformatics.rutgers.edu Department of Computer Science & BioMaPS Institute for Quantitative Biology Rutgers, The State University of New Jersey, Piscataway, NJ, USA |
|
From: Fatma M. <mil...@ya...> - 2015-05-01 22:06:45
|
Hi, In my work,i need to cluster sequencial data into groups. So, i used cygwin tocompile ghmm/tools/cluster.c using this instruction : gcc cluster.c–o c1. It works well. However when i bulit it with this instruction :./c1 test1.sqd test2.smo test3.smo, i had this message: "cluster isobsolete. If you need it rebuild the GHMM with "GHMM_OBSOLETE".Please how can i rebuild the GHMM with "GHMM_OBSOLETE? Thanks in advance. Best regards, ---- Fatma MILADI PhD Student in Computer Science Member at MIRACL laboratory |
|
From: Matthias P. <mat...@st...> - 2015-04-29 09:25:43
|
Hallo everyone, I'm rather new to GHMM and I'm having some (at least what I think) numerical problems. Here's a quick description of what I'm doing: I'm trying to train a HMM model on sequential data. The data comes from a motion capturing system. Each sample has 43 features, which are continuous. I currently do not perform any scaling and/or normalization on the data. I have attempted to scale each feature to have mean=0 and std=1, but that seems to have made things worse. Anyway, the model that I use has 10 states and I use a left-to-right architecture. I assume uniform distribution for the initial state transitions, so my transition matrix A looks something like this: A = [[0.1, 0.1, ...], [0, 0.1111, ...], ..., [0, 0, ..., 0, 1]] I always start in the first state, so my initial probabilities are: pi = [1, 0, ..., 0]. For the emission probabilities, I estimate the initial means with the K-Means algorithm using all training data. The covariance matrix is estimated using numpy.cov. I also diagonalize the covariance matrix. Here's the output of my final GHMM model when printing it: http://cl.ly/text/1C343z263f1Q (I have uploaded it as separate file to keep this somewhat readable). I'm using the latest version of GHMM with --enable-gsl and --enable-gsl-diagonal-hack set during compilation. The problems start when I attempt to train the model described above. I use baum-welch with 10 iterations and a cut-off of 0.000001. After training, the model becomes invalid and almost every vector/matrix solely consists of NaNs. Here's an example of a state 0, but all other 9 states show similar problems: http://cl.ly/text/2U3g2z0Y1I1X (again, the text file is available at the given link). Running loglikelihood before the training gives me -37434.5507181, after the training I receive nan (not surprisingly). I also sometimes receive an error message from GHMM, namely "GHMM ghmm.py:3746 - forward returned -1: Sequence 0 cannot be build.". I have no idea how I can work around the issue. Any help is greatly appreciated! If additional information is required, please also let me know! Thanks in advance – Matthias Plappert |
|
From: daehyung p. <der...@ga...> - 2015-04-01 19:44:35
|
Thanks for the reply. I appreciate about your answer. I can understand the result of viterbi is a probability density function so that it can be greater than 1. The loglikelihood is non-weighted sum of the joint distribution over the all available state path. Thus, loglikelihood(179.49) is greater than the joint distribution(171.19). There was no mention about negative loglikelihood in this library's document so that I believe these values are not negative loglikelihood. Only a remained problem is whether this value is scaled or not to avoid underflow. In this library, Baum-welch algorithm returns a scaling factor. I suspect the same factor applied to it. Is there any body know about it? Best, Daehyung |
|
From: Weipeng He <hew...@gm...> - 2015-04-01 16:19:13
|
I don't think that the score is a negative log likelihood. the result 179.49 is a log likelihood. I understand that the values are calculated in logarithmic scale to avoid underflow. But, there is no point to compute the negative of it. What i was trying to say is that the observation space is continuous, hence we use probability density function to describe the distribution (same as a Gaussian distribution or GMM). And a valid pdf can have value from zero to arbitrary large. Therefore, a valid log likelihood can be any real number. And, a joint probability of the observation sequence and any state sequence is always less than the probability of the observation sequence given any parameters. That means the output of viterbi (171.19) should be less than the log likelihood (179.49). However, if you take the value as a negative log, it will contradict the assumption. Best, Weipeng On 03/31/2015 05:53 PM, Morton, James wrote: > > I'm fairly positive that the score is a negative log likelihood. > Dealing with likelihood probabilities are not very useful due to > precision error. > > On Mar 31, 2015 9:04 AM, "Weipeng He" <hew...@gm... > <mailto:hew...@gm...>> wrote: > > Hi Daehyung, > > I don't use this library. But here is some information might solve > your question. > > The likelihood probability for continuous data is in fact a > *probability density function*, which of course can be larger than 1. > > Best regards, > Weipeng > > > On 03/31/2015 03:43 AM, daehyung park wrote: >> Hello, >> >> I am using MultivariateGaussianDistributi >> on with two dimensional sequence of float data. Now, I am really >> struggling with weird range of loglikelihood values. >> >> When I try to extract a loglikelihood from a training sequence, >> ghmm returns '179.49'. Viterbi function also returns '171.19' as >> a log value of a joint probability. Whether the sequence is a >> good or bad sample, the value should not be over than '1.0'. Am I >> wrong? >> >> Best, >> Daehyung >> >> >> ------------------------------------------------------------------------------ >> Dive into the World of Parallel Programming The Go Parallel Website, sponsored >> by Intel and developed in partnership with Slashdot Media, is your hub for all >> things parallel software development, from weekly thought leadership blogs to >> news, videos, case studies, tutorials and more. Take a look and join the >> conversation now.http://goparallel.sourceforge.net/ >> >> >> _______________________________________________ >> Ghmm-list mailing list >> Ghm...@li... <mailto:Ghm...@li...> >> https://lists.sourceforge.net/lists/listinfo/ghmm-list > > > ------------------------------------------------------------------------------ > Dive into the World of Parallel Programming The Go Parallel > Website, sponsored > by Intel and developed in partnership with Slashdot Media, is your > hub for all > things parallel software development, from weekly thought > leadership blogs to > news, videos, case studies, tutorials and more. Take a look and > join the > conversation now. http://goparallel.sourceforge.net/ > _______________________________________________ > Ghmm-list mailing list > Ghm...@li... > <mailto:Ghm...@li...> > https://lists.sourceforge.net/lists/listinfo/ghmm-list > |
|
From: Weipeng He <hew...@gm...> - 2015-03-31 15:04:35
|
Hi Daehyung, I don't use this library. But here is some information might solve your question. The likelihood probability for continuous data is in fact a *probability density function*, which of course can be larger than 1. Best regards, Weipeng On 03/31/2015 03:43 AM, daehyung park wrote: > Hello, > > I am using MultivariateGaussianDistributi > on with two dimensional sequence of float data. Now, I am really > struggling with weird range of loglikelihood values. > > When I try to extract a loglikelihood from a training sequence, ghmm > returns '179.49'. Viterbi function also returns '171.19' as a log > value of a joint probability. Whether the sequence is a good or bad > sample, the value should not be over than '1.0'. Am I wrong? > > Best, > Daehyung > > > ------------------------------------------------------------------------------ > Dive into the World of Parallel Programming The Go Parallel Website, sponsored > by Intel and developed in partnership with Slashdot Media, is your hub for all > things parallel software development, from weekly thought leadership blogs to > news, videos, case studies, tutorials and more. Take a look and join the > conversation now. http://goparallel.sourceforge.net/ > > > _______________________________________________ > Ghmm-list mailing list > Ghm...@li... > https://lists.sourceforge.net/lists/listinfo/ghmm-list |
|
From: daehyung p. <der...@ga...> - 2015-03-31 01:43:37
|
Hello, I am using MultivariateGaussianDistributi on with two dimensional sequence of float data. Now, I am really struggling with weird range of loglikelihood values. When I try to extract a loglikelihood from a training sequence, ghmm returns '179.49'. Viterbi function also returns '171.19' as a log value of a joint probability. Whether the sequence is a good or bad sample, the value should not be over than '1.0'. Am I wrong? Best, Daehyung |
|
From: Pierre D. <pie...@gm...> - 2015-01-30 04:10:29
|
Hi,
I am trying to import the ghmm python library, but i got this error :
>>> import ghmm
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File
"/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/ghmm.py",
line 112, in <module>
import ghmmwrapper
File
"/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/ghmmwrapper.py",
line 7, in <module>
import _ghmmwrapper
ImportError:
dlopen(/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/_ghmmwrapper.so,
2): Symbol not found: _RNG
Referenced from:
/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/_ghmmwrapper.so
Expected in: dynamic lookup
I already find a thread about that :
http://sourceforge.net/p/ghmm/mailman/message/25819936/
But there was no conclusive informations for me.
I had no problem with installation.
I'm running it on :
OsX 10.6.8
Swig 1.3.31
gcc 4.2.1
|
|
From: Dave dV <zed...@gm...> - 2014-12-11 06:33:16
|
Hi, I am trying to implement a model similar to a HMM profile for protein sequences (with insertion and deletion/skip states), however, I am running into some problem when training the model using Baum Welsh: > GHMM ghmm.py:148 - model.c:ghmm_dmodel_check(803): sum of s[11].out_a[*] = 0.0 (assumed final state but 2 transitions) > GHMM ghmm.py:148 - model.c:ghmm_dmodel_check(828): state 11 can't be reached but is not set as non-reachale state State 11 is one of my insertion states and it is quite natural (given the small size of my trial training set) that it would not be used during training. However, it seems that it should not be set to zero, but to a minimum non-zero value or at least epsilon, so that validation does not fail. Am I missing a parameter I can use in my HMM XML or in the arguments to the Baum Welsh function that would ensure that all defined transitions remain possible (however unlikely) and all states reachable…? Thanks in advance for your help! — Dave |
|
From: Weipeng He <hew...@gm...> - 2014-07-20 16:18:11
|
Hi,
I've installed ghmm library from yum on my Fedora 20. However, I can't
import ghmm in python.
The result is here:
>>> import ghmm
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/usr/lib64/python2.7/site-packages/ghmm.py", line 111, in <module>
import ghmmwrapper
File "/usr/lib64/python2.7/site-packages/ghmmwrapper.py", line 26, in
<module>
_ghmmwrapper = swig_import_helper()
File "/usr/lib64/python2.7/site-packages/ghmmwrapper.py", line 22, in
swig_import_helper
_mod = imp.load_module('_ghmmwrapper', fp, pathname, description)
ImportError: /usr/lib64/python2.7/site-packages/_ghmmwrapper.so:
undefined symbol:
How can I solve this?
Thank you :)
Weipeng
|
|
From: anas e. <ana...@gm...> - 2014-05-19 05:26:36
|
In trying to install GHMM, I ran into this error: /usr/bin/python setup.py build File "setup.py", line 33 SyntaxError: Non-ASCII character '\xc3' in file setup.py on line 33, but no encoding declared; see http://www.python.org/peps/pep-0263.html for details make[2]: *** [all] Error 1 This was solved by removing the german-date from the comment in line 33. I thought I should share this in case someone runs into the same problem |
|
From: anas e. <ana...@gm...> - 2014-05-16 17:52:54
|
Hi GHMM-list, I'm looking at some libraries for implementing Profile HMMs for sequence alignment problem, and I have a question about GHMM's suitability for this: is it possible in GHMM to specify a model with silent states (states that emit no signal, but could still be visited)? This is needed as it corresponds to a "deletion" occurring on the sequence. If not, do you see a way around this in GHMM? (Just having a symbol that stands for "nothing" wouldn't work because the sequences of exons we have don't contain such a symbol, so the algorithms for calculating the most likely path would wrongly exclude the silent states). Thanks, Anas |
|
From: Egon K. <ki...@gm...> - 2014-04-18 08:50:40
|
Hi, I'm writting to let you know that the lates revission apparently fails to build. I get the following error on make: .... ghmmwrapper_wrap.c:21933:3: error: too many arguments to function ‘ghmm_dmodel_cfbgibbs’ ../ghmm/cfbgibbs.h:67:7: note: declared here ghmmwrapper_wrap.c: In function ‘init_ghmmwrapper’: ghmmwrapper_wrap.c:30166:21: warning: variable ‘md’ set but not used [-Wunused-but-set-variable] error: command 'gcc' failed with exit status 1 make[2]: *** [all] Error 1 make[2]: Leaving directory `/home/ekidmose/svn-reps/ghmm/ghmmwrapper' make[1]: *** [all-recursive] Error 1 make[1]: Leaving directory `/home/ekidmose/svn-reps/ghmm' make: *** [all] Error 2 I took the lazy work-arround and reverted back to r2332, which builds fine, leading me to believe that my environment is OK. I have not checked anything in between. I just assumed the issue is with fbgibbs, which appears in logs from 2333 onwards. As a reference for anyone else facing the issue and as rusty in svn commands as me, the revert is done with this command: svn merge -r 2339:2332 . Thanks for a nice tool :) Mvh/BR Egon Kidmose |
|
From: developer.vm <dev...@gm...> - 2014-04-01 13:47:24
|
Dear Members, I am newbie to GHMM. I have recently completed my theory part of HMM and now trying my hands on actually implementing a task. I would like to know if following thing can be achieved using GHMM library. I am a Python developer with 4 years of exp so can use python binding. *Task:* To recognize spoken digit string like: 1-4-5-7-9-8 (example: a phone number). *What I have got: *Feature vectors extracted from 100 speech file each contains audio of different phone numbers. From every speech file, every 10ms I have extracted feature vector (25ms window with 10ms overlap). I know theory of Gaussian Mixture Models. If any of you guys have done something like this before and you think this can be achieved using GHMM lib (or if required with the help of scikit-learn), please help me understand HMM by practically implementing it on a task. Best Regards, Dev |
|
From: Samuel F. <se...@ca...> - 2014-03-01 21:40:02
|
I'm trying to use GHMM to write a program that will learn patterns from
music, and am having a tough time figuring out how get GHMM to understand
my data. Right now, I am just trying to create a sequence of training data
as an EmissionSequence. The code so far to do this is as follows:
c_3 = Note('c_3', 10)
d_3 = Note('f_3', 10)
e_3 = Note('e_3', 10)
f_3 = Note('f_3', 10)
g_3 = Note('g_3', 10)
a_3 = Note('a_3', 10)
b_3 = Note('b_3', 10)
c_4 = Note('c_4', 10)
alphabet = Alphabet([c_3, d_3, e_3, f_3, g_3, a_3, b_3, c_4])
alphaLen = len(alphabet)
tran_seq = EmissionSequence(alphabet, [c_3, d_3])
Where my Note class is:
class Note:
def __init__(self, pitch, duration):
self.pitch = pitch
self.duration = duration
def __str__(self):
return "Pitch: {} Duration: {}\n".format(self.pitch,
self.duration)
def __len__(self):
return 1
When I try to run the my program, I get the following error:
Exception AttributeError: "'EmissionSequence' object has no attribute
'cseq'" in <bound method EmissionSequence.__del__ of <ghmm.EmissionSequence
object at 0x7f3d43816450>> ignored
I think my problem is that I'm not entirely understanding the constructor
for EmissionSequence. I've take a look through the email archives, and at
the examples in the GHMM tests, but haven't been able to find any examples
with an EmissionSequence being created out of a set of objects.
Is what I'm trying to do possible, or do I need to encode my data as
strings in order to create an Emission Sequence out of it?
Thank,
Sam
|
|
From: <bru...@gm...> - 2014-01-30 05:10:43
|
It could be that your variance or transitions are to small since the probabilities are overflowing to zero during the forward algorithm. Do you have outliers like 600 or 50 in your data assuming your mean still is 300 and variance is 50. Try increasing the variance to see if it solves the problem. You could also create another state for noise or remove outliers before you train.
On Jan 29, 2014, at 1:10 AM, Alexey Morozov <ale...@gm...> wrote:
> verbose.Str() tells me that transition probabilities are defined correctly, but Baum-Welch of Viterbi still refuse to run, returning
>
> NO sequence can be build from smodel smo!
> GHMM ghmm.py:148 - sreestimate.c:ghmm_cmodel_baum_welch(878): sreestimate_one_step false (1.step)
>
> and
>
> GHMM ghmm.py:148 - sviterbi.c:ghmm_cmodel_viterbi(215): sequence can't be build from model
> ([], -inf)
>
> respectively.
> I've changed constant you mentioned to 1e-100, but it didn't help. I've even set it equal to 10e-1000, zero (or values so low that they overflow double-precision and become zero during compilation), nothing changed. At the same time some toy HMMs with sequences of tens of values work correctly (both Baum-Welch and Viterbi).
>
> I've also changed GHMM_MAX_SEQ_LEN to fit my long sequences since library seems tobreak after a certain size. Still no effect.
>
> 2014-01-28 eric brugel <bru...@gm...>
> I believe print model rounds the transition probabilities to two decimal places. Use print model.verboseStr() to get the true values. When Buam-Welch is run if the forward algorithm probabilities is less than a constant GHMM_EPS_PREC defined in the ghmm.h the algorithm breaks. You could change this value and recompile. I think it is 10e-4 right now.
>
>
> On Tue, Jan 28, 2014 at 2:27 AM, Alexey Morozov <ale...@gm...> wrote:
> Is there any lower limit for transition values? I've tried to create an HMM via following code:
>
> print('Creating HMM object...\n')
> f=ghmm.Float()#Emission domain
> A=[[0.999999,0.000001],[0.000001,0.999999]]#Transition matrix
> B=[[300,50],[-300,50]] #Emission probabilities
> pi=[0.5,0.5] #Starting state can be any with equal probability
> model=ghmm.HMMFromMatrices(f,ghmm.GaussianDistribution(f),A,B,pi)
>
> so that I'd have long streaks of either state, changing once in a megabase or so (parameter in question in real genomes comes in positive/negative blocks from several hundreds of kilobases to several megabases long). So, add
>
> print A
> print model
>
> It returns
>
> [[0.999999, 1e-06], [1e-06, 0.999999]]
> GaussianEmissionHMM(N=2)
> state 0 (initial=0.50, mu=300.00, sigma=50.00)
> Transitions: ->0 (1.00), ->1 (0.00)
> state 1 (initial=0.50, mu=-300.00, sigma=50.00)
> Transitions: ->0 (0.00), ->1 (1.00)
>
> So my code gives GHMM 10**-6 and GHMM thinks it's 0. Baum-Welch refuses to run and model that never changes states is useless anyway. How do I fix such behaviour?
> --
> Alexey Morozov,
> LIN SB RAS, bioinformatics group.
> Irkutsk, Russia.
>
> ------------------------------------------------------------------------------
> WatchGuard Dimension instantly turns raw network data into actionable
> security intelligence. It gives you real-time visual feedback on key
> security issues and trends. Skip the complicated setup - simply import
> a virtual appliance and go from zero to informed in seconds.
> http://pubads.g.doubleclick.net/gampad/clk?id=123612991&iu=/4140/ostg.clktrk
> _______________________________________________
> Ghmm-list mailing list
> Ghm...@li...
> https://lists.sourceforge.net/lists/listinfo/ghmm-list
>
>
>
>
>
> --
> Alexey Morozov,
> LIN SB RAS, bioinformatics group.
> Irkutsk, Russia.
|
|
From: Alexey M. <ale...@gm...> - 2014-01-29 06:10:27
|
verbose.Str() tells me that transition probabilities are defined correctly,
but Baum-Welch of Viterbi still refuse to run, returning
NO sequence can be build from smodel smo!
GHMM ghmm.py:148 - sreestimate.c:ghmm_cmodel_baum_welch(878):
sreestimate_one_step false (1.step)
and
GHMM ghmm.py:148 - sviterbi.c:ghmm_cmodel_viterbi(215): sequence can't be
build from model
([], -inf)
respectively.
I've changed constant you mentioned to 1e-100, but it didn't help. I've
even set it equal to 10e-1000, zero (or values so low that they overflow
double-precision and become zero during compilation), nothing changed. At
the same time some toy HMMs with sequences of tens of values work correctly
(both Baum-Welch and Viterbi).
I've also changed GHMM_MAX_SEQ_LEN to fit my long sequences since library
seems to break after a certain size. Still no effect.
2014-01-28 eric brugel <bru...@gm...>
> I believe print model rounds the transition probabilities to two decimal
> places. Use print model.verboseStr() to get the true values. When
> Buam-Welch is run if the forward algorithm probabilities is less than a
> constant GHMM_EPS_PREC defined in the ghmm.h the algorithm breaks. You
> could change this value and recompile. I think it is 10e-4 right now.
>
>
> On Tue, Jan 28, 2014 at 2:27 AM, Alexey Morozov <
> ale...@gm...> wrote:
>
>> Is there any lower limit for transition values? I've tried to create an
>> HMM via following code:
>>
>> print('Creating HMM object...\n')
>> f=ghmm.Float()#Emission domain
>> A=[[0.999999,0.000001],[0.000001,0.999999]]#Transition matrix
>> B=[[300,50],[-300,50]] #Emission probabilities
>> pi=[0.5,0.5] #Starting state can be any with equal probability
>> model=ghmm.HMMFromMatrices(f,ghmm.GaussianDistribution(f),A,B,pi)
>>
>> so that I'd have long streaks of either state, changing once in a
>> megabase or so (parameter in question in real genomes comes in
>> positive/negative blocks from several hundreds of kilobases to several
>> megabases long). So, add
>>
>> print A
>> print model
>>
>> It returns
>>
>> [[0.999999, 1e-06], [1e-06, 0.999999]]
>> GaussianEmissionHMM(N=2)
>> state 0 (initial=0.50, mu=300.00, sigma=50.00)
>> Transitions: ->0 (1.00), ->1 (0.00)
>> state 1 (initial=0.50, mu=-300.00, sigma=50.00)
>> Transitions: ->0 (0.00), ->1 (1.00)
>>
>> So my code gives GHMM 10**-6 and GHMM thinks it's 0. Baum-Welch refuses
>> to run and model that never changes states is useless anyway. How do I fix
>> such behaviour?
>> --
>> Alexey Morozov,
>> LIN SB RAS, bioinformatics group.
>> Irkutsk, Russia.
>>
>>
>> ------------------------------------------------------------------------------
>> WatchGuard Dimension instantly turns raw network data into actionable
>> security intelligence. It gives you real-time visual feedback on key
>> security issues and trends. Skip the complicated setup - simply import
>> a virtual appliance and go from zero to informed in seconds.
>>
>> http://pubads.g.doubleclick.net/gampad/clk?id=123612991&iu=/4140/ostg.clktrk
>> _______________________________________________
>> Ghmm-list mailing list
>> Ghm...@li...
>> https://lists.sourceforge.net/lists/listinfo/ghmm-list
>>
>>
>
--
Alexey Morozov,
LIN SB RAS, bioinformatics group.
Irkutsk, Russia.
|
|
From: eric b. <bru...@gm...> - 2014-01-28 14:19:22
|
I believe print model rounds the transition probabilities to two decimal
places. Use print model.verboseStr() to get the true values. When
Buam-Welch is run if the forward algorithm probabilities is less than a
constant GHMM_EPS_PREC defined in the ghmm.h the algorithm breaks. You
could change this value and recompile. I think it is 10e-4 right now.
On Tue, Jan 28, 2014 at 2:27 AM, Alexey Morozov <ale...@gm...
> wrote:
> Is there any lower limit for transition values? I've tried to create an
> HMM via following code:
>
> print('Creating HMM object...\n')
> f=ghmm.Float()#Emission domain
> A=[[0.999999,0.000001],[0.000001,0.999999]]#Transition matrix
> B=[[300,50],[-300,50]] #Emission probabilities
> pi=[0.5,0.5] #Starting state can be any with equal probability
> model=ghmm.HMMFromMatrices(f,ghmm.GaussianDistribution(f),A,B,pi)
>
> so that I'd have long streaks of either state, changing once in a megabase
> or so (parameter in question in real genomes comes in positive/negative
> blocks from several hundreds of kilobases to several megabases long). So,
> add
>
> print A
> print model
>
> It returns
>
> [[0.999999, 1e-06], [1e-06, 0.999999]]
> GaussianEmissionHMM(N=2)
> state 0 (initial=0.50, mu=300.00, sigma=50.00)
> Transitions: ->0 (1.00), ->1 (0.00)
> state 1 (initial=0.50, mu=-300.00, sigma=50.00)
> Transitions: ->0 (0.00), ->1 (1.00)
>
> So my code gives GHMM 10**-6 and GHMM thinks it's 0. Baum-Welch refuses to
> run and model that never changes states is useless anyway. How do I fix
> such behaviour?
> --
> Alexey Morozov,
> LIN SB RAS, bioinformatics group.
> Irkutsk, Russia.
>
>
> ------------------------------------------------------------------------------
> WatchGuard Dimension instantly turns raw network data into actionable
> security intelligence. It gives you real-time visual feedback on key
> security issues and trends. Skip the complicated setup - simply import
> a virtual appliance and go from zero to informed in seconds.
>
> http://pubads.g.doubleclick.net/gampad/clk?id=123612991&iu=/4140/ostg.clktrk
> _______________________________________________
> Ghmm-list mailing list
> Ghm...@li...
> https://lists.sourceforge.net/lists/listinfo/ghmm-list
>
>
|
|
From: Alexey M. <ale...@gm...> - 2014-01-28 07:27:52
|
Is there any lower limit for transition values? I've tried to create an HMM
via following code:
print('Creating HMM object...\n')
f=ghmm.Float()#Emission domain
A=[[0.999999,0.000001],[0.000001,0.999999]]#Transition matrix
B=[[300,50],[-300,50]] #Emission probabilities
pi=[0.5,0.5] #Starting state can be any with equal probability
model=ghmm.HMMFromMatrices(f,ghmm.GaussianDistribution(f),A,B,pi)
so that I'd have long streaks of either state, changing once in a megabase
or so (parameter in question in real genomes comes in positive/negative
blocks from several hundreds of kilobases to several megabases long). So,
add
print A
print model
It returns
[[0.999999, 1e-06], [1e-06, 0.999999]]
GaussianEmissionHMM(N=2)
state 0 (initial=0.50, mu=300.00, sigma=50.00)
Transitions: ->0 (1.00), ->1 (0.00)
state 1 (initial=0.50, mu=-300.00, sigma=50.00)
Transitions: ->0 (0.00), ->1 (1.00)
So my code gives GHMM 10**-6 and GHMM thinks it's 0. Baum-Welch refuses to
run and model that never changes states is useless anyway. How do I fix
such behaviour?
--
Alexey Morozov,
LIN SB RAS, bioinformatics group.
Irkutsk, Russia.
|
|
From: R. B. S. <bur...@gm...> - 2013-10-19 14:57:24
|
Hi all, I am trying to create an profile HMM with match, insertion and deletion transition states similar to figure 1 here: http://nar.oxfordjournals.org/content/41/10/e109.full I am not sure if I have to create the trains ion states nor how I would do that considering there are four nucleotide residues that contribute to a match. Does anyone have an example of a profile HMM with the match, deletion and insertions states or know where I can find one coded in GHMM? Thanks, Burke |
|
From: Anisa A. <ani...@gm...> - 2013-08-26 15:56:01
|
Hi Eric, Thanks for the paper. I've read it thoroughly and it was mentioned in page 394 that equation 5 ensures that this distance is non-negative. Can you please clarify this? My other question is regarding the *sequence length* in the distance function. It seems that a number of observations is needed for the distance method to be converged and this number is dependent on N and M. I'd like to know if we should check this convergence in our own code or it is already handled. Is it possible that it doesn't converge at all? Please let me know if there exist more resources for the implementation of the distance method so I can study more on details. Many thanks, Anisa On Thu, Aug 22, 2013 at 3:20 PM, eric brugel <bru...@gm...> wrote: > http://cronos.rutgers.edu/~lrr/publications.html #227. It is defined as a > difference of logs so it could be negative. The asymmetric distance is > implemented. > > > On Thu, Aug 22, 2013 at 7:29 AM, Anisa Allahdadi <ani...@gm...>wrote: > >> Hi, >> >> Regarding the following thread: >> https://sourceforge.net/p/ghmm/mailman/message/27217546/ >> >> I'd like to know whether the distance should be symmetric or no. ( >> distance( model0, model1 ) =? distance( model1, model0 ) ) >> >> I'm still trying to find the book mentioned in the previous thread >> explaining how the distance method works, but as I worked with the distance >> function, I got irrational results, such as negative distance or asymmetric >> distance! >> >> Any hint, clue or example is very appreciated. >> >> Thanks in advance, >> Anisa >> >> >> ------------------------------------------------------------------------------ >> Introducing Performance Central, a new site from SourceForge and >> AppDynamics. Performance Central is your source for news, insights, >> analysis and resources for efficient Application Performance Management. >> Visit us today! >> >> http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk >> _______________________________________________ >> Ghmm-list mailing list >> Ghm...@li... >> https://lists.sourceforge.net/lists/listinfo/ghmm-list >> >> > |
9 messages has been excluded from this view by a project administrator.