Supervised Learning question & numpy issue

[olschwab]

Hello,

First, I wanted to point out that the current version of numpy (1.12.0) raises the following issue when using your code: rbgirshick/py-faster-rcnn#481. I used numpy 1.11.0 for the time being.

For my application, I would like to test your implementation of HSMM, however I have a supervised classification problem. Does your code also contain methods for supervised training?

[Matthew Johnson]

Thanks for letting us know about the float index issue! Can you tell us what line it happens on and/or provide an example that causes the error?

For supervised training, it depends on what your supervised training objective is and what optimization algorithm you intend to use. If you can set up your problem so that all you need from the HSMM part is to be able to draw samples or compute some particular expectations and not differentiate through those computations, then you can probably use parts of this library. If you want to differentiate an HSMM log likelihood then this the E-step code in this library can also help, though it's a bit tricky to set up. However, this library doesn't support differentiating through the expected sufficient statistics or sampling procedures.

[olschwab]

Actually, the numpy error seems to come from pybasicbayes (line 224 in sample_invwishart: x = np.random.randn(nu,n)).

My problem is a segmentation/classification task. I basically hope to infer aircraft configuration (flaps and gears setting) from positional data. Put differently I have events described by z(t), v(t) and I want to figure out flaps (=1,2,3,4) and gears (=0,1) settings per timestep.
I tried experimenting with scikit-learn, considering each timestep as independent event. However I do think the dynamics are important and I need to consider sequences. This is because 1) I can have significant noise (wind, sideslip, atmosphere) and 2) because my observational variables do not lead to separable classification outputs (e.g. the same height and velocity can be achieved in different flap settings if the angle of attack, unknown to me, is changed). This is what lead me to H(S)MMs, hoping that considering sequences could improve my results.