Hello all
I saw that HMMs will be removed in version 0.17. As a lover of HMMs and
sklearn, in an attempt to save them, a friend and myself have been working
on a cython-optimized HMM package which we think may be appropriate for
sklearn.
The repo is here: https://github.com/jmschrei/yahmm
To summarize, it implements forward, backward, forward-backward, viterbi,
baum-welch training, and viterbi training. It allows for silent states,
normalizes out-edges to sum to a probability of 1., and will try to
simplify your graph structure. Forward, backward, and viterbi are
implemented in O(n*m) time instead of O(m^2), where n is the average number
of edges per state and m is the number of states.
A model is can contain states with different distribution types, instead of
being limited to each character-generating state being of the same
distribution type. Currently many distributions and kernel densities are
implemented, but the user can make their own arbitrary distribution on the
fly and have it work with this package.
The downsides are that it's hard to use its most expressive features in the
current sklearn framework of defining a classifier in one line (using
Model.from_matrix() ). Instead, it may take several lines to define the
underlying model in our package.
I'd love to hear feedback from others on the idea of merging this with
sklearn.
------------------------------------------------------------------------------
"Accelerate Dev Cycles with Automated Cross-Browser Testing - For FREE
Instantly run your Selenium tests across 300+ browser/OS combos. Get
unparalleled scalability from the best Selenium testing platform available.
Simple to use. Nothing to install. Get started now for free."
http://p.sf.net/sfu/SauceLabs
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
