Open eliotmoss opened 7 years ago
Glad that it's running now! Perhaps the easier way to get this is to look in model.states_list
, which contains a list of _SLDSStates
objects, one for each dataset you've added to the model. Inside the states object you'll find the following fields:
stateseq
: the discrete state for each timegaussian_states
: the continuous states for each timeAs
, Bs
, Cs
, Ds
, etc : stacks of dynamics matrices used at each point in time. I realize this is a bit indirect. If you look at the states code (https://github.com/mattjj/pyslds/blob/master/pyslds/states.py#L131) you can see how to get the dynamics for each discrete latent state. We should add these dynamics properties directly to the model. That would make this easier.
On 5/14/2017 10:25 AM, Scott Linderman wrote:
Thank you, Scott, that should get me farther.
Do you have any plotting examples, such as what was used to make the SLDS (not rSLDS) figures in the rSLDS paper? I am thinking of the Lorenz attractor or Synthetic NASCAR figures. I realize they were chosen precisely because rSLDS does notably better than SLDS on those examples, but seeing how you get the figures would help me :-) ...
Hi Eliot, I do have those examples, but I'm tied up with NIPS submissions this week. I'll get them to you next week though! Best, Scott
On Sun, May 14, 2017 at 2:48 PM, Eliot Moss notifications@github.com wrote:
On 5/14/2017 10:25 AM, Scott Linderman wrote:
Thank you, Scott, that should get me farther.
Do you have any plotting examples, such as what was used to make the SLDS (not rSLDS) figures in the rSLDS paper? I am thinking of the Lorenz attractor or Synthetic NASCAR figures. I realize they were chosen precisely because rSLDS does notably better than SLDS on those examples, but seeing how you get the figures would help me :-) ...
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/mattjj/pyslds/issues/13#issuecomment-301331842, or mute the thread https://github.com/notifications/unsubscribe-auth/AFXwKPcy2BP1_et587uSBshk8Fp5GmH4ks5r50yMgaJpZM4NaQTI .
On 5/16/2017 3:11 PM, Scott Linderman wrote:
Hi Eliot, I do have those examples, but I'm tied up with NIPS submissions this week. I'll get them to you next week though!
Sure - I understand. Good luck with NIPS! E
Thank you for the assistance figuring out installation and building. I can run the pyslds examples under python2 now (under python3 they run, but the graphics don't pop up).
I am able to take one of my time series, build a model, and plot observations and smoothed observations. Yay!
What I have not been able to figure out is how to get the trained A, B, etc., matrices out, or how to make a display similar to (say) the evolution of the Lorentz attractor or a display of what is happening in phase space (the latent dimensions).
Do you have examples of how to get at that information and to plot it? Seeing phase space evolution with colored dots (different colors for different states) would help me see what my models are extracting from the data, etc. As it stands I don't understand your layers of code well enough to do this ...