[MRG] Multithreaded HMM training

[jmschrei]

In progress

[jmschrei]

This is almost ready. Just need to add benchmarks, add more thorough unit tests, and we should be good to go.

[jmschrei]

Preliminary single core results:

master

HIDDEN MARKOV MODEL BENCHMARKS
gaussian emissions
FORWARD         : time: 11.998, logp: -62.823
BACKWARD        : time: 14.207, logp: -62.823
VITERBI         : time: 9.1404, logp: -64.18
FORWARD-BACKWARD: time: 34.526
BW TRAINING     : time: 4.9141, improvement: 4495.0

multivariate gaussian emissions
FORWARD         : time: 27.994, logp: -922.76
BACKWARD        : time: 28.35, logp: -922.76
VITERBI         : time: 23.949, logp: -922.76
FORWARD-BACKWARD: time: 77.954
BW TRAINING     : time: 2.3619, improvement: 8.7799e+04

branch

HIDDEN MARKOV MODEL BENCHMARKS
gaussian emissions
FORWARD         : time: 4.9733, logp: -62.823
BACKWARD        : time: 5.0479, logp: -62.823
VITERBI         : time: 8.3918, logp: -64.18
FORWARD-BACKWARD: time: 17.787
BW TRAINING     : time: 1.3614, improvement: 4495.0

multivariate gaussian emissions
FORWARD         : time: 5.474, logp: -922.76
BACKWARD        : time: 5.6663, logp: -922.76
VITERBI         : time: 23.631, logp: -922.76
FORWARD-BACKWARD: time: 33.408
BW TRAINING     : time: 0.3183, improvement: 8.7799e+04

Only forward, backward, and training have been optimized, so it makes sense that forward-backward and Viterbi don't see huge improvements (forward-backward calls both forward and backward, so it is sped up a little). Master branch includes the speed improvements gained in the GIL released distribution section; these improvements are based only on improvements to the HMM code.

[jmschrei]

Multithreading training currently helps a lot with big models, but can be harmful for small models. This is because I'm calling the GIL a lot to ensure all the data structures are threadsafe. This model has 151 states.

HIDDEN MARKOV MODEL BENCHMARKS
gaussian emissions
BW TRAINING (1 thread) : time: 4.2026, improvement: 3846.1
BW TRAINING (4 threads): time: 2.2716, improvement: 3846.1

multivariate gaussian emissions
BW TRAINING (1 thread) : time: 8.8531, improvement: 7.3234e+05
BW TRAINING (4 threads): time: 4.6516, improvement: 7.3234e+05

In contrast, this model has 16 states.

HIDDEN MARKOV MODEL BENCHMARKS
gaussian emissions
BW TRAINING (1 thread) : time: 0.063518, improvement: 587.19
BW TRAINING (4 threads): time: 0.50892, improvement: 587.19

multivariate gaussian emissions
BW TRAINING (1 thread) : time: 0.15548, improvement: 7.1882e+04
BW TRAINING (4 threads): time: 1.2511, improvement: 7.1882e+04

[jmschrei]

branch

discrete distribution
FORWARD         : time: 8.1359, logp: -30.923
BACKWARD        : time: 8.2184, logp: -30.923
VITERBI         : time: 12.186, logp: -37.664
FORWARD-BACKWARD: time: 28.234
BW TRAINING     : time: 1.4654, improvement: 697.88

master

discrete distribution
FORWARD         : time: 16.173, logp: -30.923
BACKWARD        : time: 18.447, logp: -30.923
VITERBI         : time: 12.245, logp: -37.664
FORWARD-BACKWARD: time: 45.645
BW TRAINING     : time: 4.4667, improvement: 697.88

[jmschrei]

The only issue is one data structure is not thread-safe. I am currently looking for it.

[jmschrei]

Everything works, finally. Merging!