TODO

Try to make both matlab + python code parallelizable

Figure out why numpy linalg solve is so slow compared to matlab.  
Multithreaded? No, it seems that numpy is very naive about triangular solvers


CONFIGURATION FILES -------------
How to cleanly store the parameters for a prior distribution so that they can be easily checked/modified/etc.

IDEA 0: Flat text file

MixModel.config
--------
alpha0=1
alpha1=5.0
K=25
invW = eye(7)

PROS: simple and easy to implement
CONS: cannot easily init "from data" ... e.g. set invW = eye(D) 

IDEA 1:  Use a python script!

MixModelPriorConfig.py
   must expose either:
       global var "foo"
       or
       function called "set_foo"  that takes Data object and cmdline args

PROS: 
CONS: complicated to understand, liable to get design wrong, tougher to override at cmdline

ALLOCATION MODEL ---------------

Change GroupIDs so it uses only nGroup+1 space, not 2*nGroup

BENCHMARKS -------------
Compare EM for GMM with fancy objects to plain single file impl.

SLOW simple version  ~ 27 sec / iter
EM for Mixture of 25 Gaussians | seed=6114
    1/3 after 27 sec | 2.54015013e+07
    2/3 after 54 sec | 3.02531705e+07
    3/3 after 82 sec | 3.38816615e+07
    
FAST simple version (dotABT instead of np.dot, etc)   ~21 sec / iteration
EM for Mixture of 25 Gaussians | seed=6114
    1/2 after 21 sec | 2.54015013e+07
    2/2 after 42 sec | 3.02531705e+07
    
Fancy version BEFORE improvments:   ~28 sec per iter (first iter has extra work for init)
      0/3 after     36 sec. |  evidence  2.520837648e+07
      1/3 after     62 sec. |  evidence  2.757578112e+07
      
Fancy version AFTER improvements (dotABT, etc)...  ~23 sec per iteration. how to get more?
      0/4 after     33 sec. |  evidence  2.520795538e+07
      1/4 after     55 sec. |  evidence  2.716296063e+07
      2/4 after     79 sec. |  evidence  3.343968874e+07
      3/4 after    102 sec. |  evidence  3.590508356e+07