Api fixes and tests

[franknoe]

Large PR with a number of fixes and improvements. All tests pass on my machine, and the examples look fine. In order to generate output (both figures and tables) go to an example directory, e.g. `cd examples/p5ab-hairpin/', and do:

• python generate-figure.py generates the three state-assignment figures as pdf. Set --nstates or --nsample or --verbose if you want.
• pdflatex doc generates the LaTeX table.

Major changes in the code:

New features:

• Offering useful functions in a user API (bhmm/api.py). The old api objects (HMM, BHMM, MLHMM) are still available.
• Extended HMM to compute various useful properties.
• Extended MLHMM (MaximumLikelihoodEstimator) to provide various useful properties.
• New SampledHMM object that encapsulates a list of sampled HMMs and can provide various statistical quantities (means, stds, confidence intervals)
• New HMM subclasses to provide convenient access to special HMM types (Gaussian, Discrete)

Testing and docs:

• Added extensive Tests for MLHMM and BHMM
• Removed verbosity in tests
• Improved API docs

Refactoring:

• HMM now has an update option to set new transition matrix. Avoids inconsistencies and removes inconsistency errors previous created by directly setting attributes
• HMM, BHMM and MLHMM have all their attributes hidden, and have property getters to avoid undesired write access.
• Moved general settings (kernel, dtype and verbose) to bhmm.config and set to reasonable defaults
• Refactored various internal implementation details and moved all python files into packages.
• Moved plots module to examples directory and thus out of the install directory. Consequently removed the dependencies to matplotlib and seaborn.

Examples for paper

• Made example scripts more general and simpler
• Fixed LaTeX table generation for BHMM data
• Removed generated files from git

Bugfixes

• Fixed: initialization problems with discrete and Gaussian HMMs
• Fixed: Using pyEMMA>=1.2, should fix previous dependency problems
• Countless small bugfixes

[jchodera]

This is an enormous PR, but looks very nice! Many thanks for putting this together.

In the future (when both of us have more availability for rapid code review!), let's try to make PRs more atomic and associated with single features. But this is totally fine for now.

I'd love to discuss where the plotting and data analysis tools should actually go. If they're meant to be general command-line scripts, we can actually move them into bhmm.scripts and deploy them on install into the user's path. Also, I do think it is very useful to have the general plotting tools installed, rather than keeping them in the examples directory---especially if they are to be used frequently. But we can discuss this in separate issues.

Merging!

[franknoe]

Am 05/05/15 um 03:22 schrieb John Chodera:

This is an enormous PR, but looks very nice! Many thanks for putting this together.

In the future (when both of us have more availability for rapid code review!), let's try to make PRs more atomic and associated with single features. But this is totally fine for now.

I totally agree. But this was actually just two days work - at times when you are available I can of course still subdivide it!

I'd love to discuss where the plotting and data analysis tools should actually go. If they're meant to be general command-line scripts, we can actually move them into |bhmm.scripts| and deploy them on install into the user's path. Also, I /do/ think it is very useful to have the general plotting tools installed, rather than keeping them in the examples directory---especially if they are to be used frequently. But we can discuss this in separate issues.

I agree but would like to avoid a hard dependency. We could have them as an optional dependency, check for their availability where we need them and raise an Error when they are not there. Alternatively, we could build a user package which depends on BHMM and everything else that may be useful and adds convenience functions, plotting functions and stuff for specific HMMs.

Best Frank

Merging!

— Reply to this email directly or view it on GitHub https://github.com/bhmm/bhmm/pull/17#issuecomment-98899182.

---

Prof. Dr. Frank Noe Head of Computational Molecular Biology group Freie Universitaet Berlin

Phone: (+49) (0)30 838 75354 Web: research.franknoe.de

Mail: Arnimallee 6, 14195 Berlin, Germany