HSMM-SLDS

[vruetten]

Hi Matt and Scott,

Thanks for the awesome toolbox.

I’ve been having some trouble when trying to link a HSMM and an SLDS.
A “NotImplementedError” error gets raised when I try to generate from the model. (The HMMSLDS version works fine.) I'm not sure if I'm doing something wrong.

For instance if I try to generate from a HSMMSLDS, I get the following error.
Thanks for letting me know what I should do (and again for making your code open-source)

import pyslds.models as mm  
import numpy as np
from pyhsmm.basic import distributions
D_latent = 2                  
D_obs = 2                        
D_input = 0        
K = 2

mu_init = np.zeros(D_latent)
mu_init[0] = 1.0
sigma_init = 0.01 * np.eye(D_latent)

alpha_0,beta_0 = 20.0, 2.0
betas = [(alpha_0,beta_0) for state in range(K)]
As = [np.eye(D_latent) for state in range(K)]
Cs = [np.random.randn(D_obs, D_latent) for state in range(K)]
init_dynamics_distns = [distributions.Gaussian(mu=mu_init, sigma=sigma_init) for _ in range(K)]
dynamics_distns = [distributions.Regression(A=A, sigma=0.01*np.eye(D_latent)) for A in As]
emission_distns = [distributions.Regression(A=C, sigma=0.01*np.eye(D_latent)) for C in Cs]
dur_distns=[distributions.PoissonDuration(alpha_0=alpha,beta_0=beta)
                       for (alpha, beta) in betas]
    
model=mm.HSMMSLDS(
       init_dynamics_distns=init_dynamics_distns,
       dynamics_distns=dynamics_distns,
       emission_distns=emission_distns,
       dur_distns=dur_distns,
       init_state_distn='uniform',
       alpha=3.0)

model.generate(1)

-----------------------------------
AttributeErrorTraceback (most recent call last)
<ipython-input-6-1ab324cecf4b> in <module>()
     29         alpha=3.0)
     30 
---> 31 model.generate(1)

~/anaconda/lib/python3.5/site-packages/pyslds/models.py in generate(self, T, keep, **kwargs)
     37 
     38     def generate(self, T=100, keep=True, **kwargs):
---> 39         s = self._states_class(model=self, T=T, initialize_from_prior=True, **kwargs)
     40         s.generate_states()
     41         data = self._generate_obs(s)

~/anaconda/lib/python3.5/site-packages/pyslds/states.py in __init__(self, model, data, mask, fixed_stateseq, **kwargs)
    754         super(_SLDSStatesCountData, self). \
    755             __init__(model, data=data, mask=mask,
--> 756                      fixed_stateseq=fixed_stateseq, **kwargs)
    757 
    758         # Check if the emission matrix is a count regression

~/anaconda/lib/python3.5/site-packages/pyslds/states.py in __init__(self, model, T, data, inputs, mask, stateseq, gaussian_states, generate, initialize_from_prior, fixed_stateseq)
    553                                                     generate=generate,
    554                                                     initialize_from_prior=initialize_from_prior,
--> 555                                                     fixed_stateseq=fixed_stateseq)
    556 
    557     @property

~/anaconda/lib/python3.5/site-packages/pyslds/states.py in __init__(self, model, T, data, inputs, stateseq, gaussian_states, generate, initialize_from_prior, fixed_stateseq)
     47                 self.resample()
     48             else:
---> 49                 self.generate_states()
     50 
     51     def generate_states(self):

~/anaconda/lib/python3.5/site-packages/pyslds/states.py in generate_states(self)
     50 
     51     def generate_states(self):
---> 52         super(_SLDSStates,self).generate_states()
     53         self.generate_gaussian_states()
     54 

~/anaconda/lib/python3.5/site-packages/pyhsmm/internals/hsmm_states.py in generate_states(self)
    113     # potentials?
    114     def generate_states(self):
--> 115         if self.left_censoring:
    116             raise NotImplementedError
    117         idx = 0

AttributeError: 'HSMMSLDSStatesEigen' object has no attribute 'left_censoring'

[zqwei]

@vrutten I added some details for installation. Hope it helps. https://github.com/zqwei/pyslds

[slinderman]

Thanks Ziqiang. @mattjj, can you update pip with the latest pylds so that we don't have to install it from source? In the meantime, I'll remove the hips dependency.

…

On Mon, Feb 26, 2018 at 9:59 AM, Ziqiang Wei ***@***.***> wrote: @vrutten <https://github.com/vrutten> I added some details for installation. Hope it helps. https://github.com/zqwei/pyslds — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#25 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AFXwKPSPa59QY9BEHpT2kmC-bEWJzdilks5tYsbBgaJpZM4STKeX> .

[vruetten]

Hi Ziqiang(!)
Funny to find you through this interface.

I'm not quite sure why you reference "details for installation.".
Installation/models work fine. I just seem to have issues with the HSMMSLSD one.
Thanks for any clarifications.

[slinderman]

Ah, good point. The HSMM + SLDS is a relatively untested combination. I'll look into the generation code and see if I can debug it. Best, Scott

…

On Mon, Feb 26, 2018 at 11:25 AM, vmsr2 ***@***.***> wrote: Hi Ziqiang(!) Funny to find you through this interphase. I'm not quite sure why you reference "details for installation.". Installation/models work fine. I just seem to have issues with the HSMMSLSD one. Thanks for any clarifications. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#25 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AFXwKFp7i75YLqSAST7FxZwR_I9EV455ks5tYtsUgaJpZM4STKeX> .

[zqwei]

@vrutten Hi, Virginia, as Scott mentioned, some of the code uses updated libs and I wrote a summary which should be pip-installed with independent effort to make the code runnable (if you did not realize that).

[vruetten]

Hi again,

Would you have a maths write up of the derivations associated with the various models?
It would make following the code quite a bit easier. (Or point me to a paper which has them explicitly.)

Thanks!

[vruetten]

Hi again :)
Double checking (in order not to duplicate work)
Gibbs sampling works fine, VBEM works for the DefaultSLDS model, but not for all the other models (e.g.: HSMMSLDS).
I get errors when trying to run inference using the Viterbi approximation on the DefaultSLDS. Could you tell me if the code is supposed to be working (i.e.: there's a simple implementation issue on my half) or if that part of the code is still to be debugged?

thanks