markovmodel / PyEMMA / 1697 / 1
87%
devel: 87%

DEFAULT BRANCH: devel
Ran
Files 156
Run time 10min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
coverage: 84.642% (-2.5%) from 87.119%
python=2.7 CONDA_PY=27 CONDA_NPY=19

push

travis-ci

marscher 
Merge pull request #694 from markovmodel/merge_hmm_in_devel

Brings significantly improved Hidden Markov Model code. This is the PyEMMA interface for bhmm 0.6. Features:
* Maximum likelihood estimation can deal with disconnected hidden transition matrices. The desired connectivity is selected only at the end of the estimation (optionally), or a posteriori.
* Much more robust estimation of initial Hidden Markov model. 
* Added option stationary that controls whether input data is assumed to be sampled from the stationary distribution (and then the initial HMM distribution is taked as the stationary distribution of the hidden transition matrix), or not (then it's independently estimated using the EM standard approach). Default: stationary=False. This *changes the default behavior* w.r.t. the previous version, but in a good way: Now the maximum-likelihood estimator always converges. Unfortunately that also means it is much slower compared to previous versions which stopped without proper convergence.
* Hidden connectivity: By default delivers a HMM with the full hidden transition matrix, that may be disconnected. This *changes the default behavior* w.r.t. the previous version. Set connectivity='largest' or connectivity='populous' to focus the model on the largest or most populous connected set of hidden states
* Provides a way to measure connectivity in HMM transition matrices: A transition only counts as real if the hidden count matrix element is larger than mincount_connectivity (by default 1 over the number of hidden states). This seems to be a much more robust metric of real connectivity than MSM count matrix connectivity.
* Observable set: If HMMs are used for MSM coarse-graining, the MSM active set will become the observed set (as before). If a HMM is estimated directly, by default will focus on the nonempty set (states with nonzero counts in the lagged trajectories). Optionally can also use the full set labels - in this case no indexing or r... (continued)

11127 of 13146 relevant lines covered (84.64%)

0.85 hits per line

Source Files on job 1697.1 (python=2.7 CONDA_PY=27 CONDA_NPY=19)