b56759009dbeba4a32f124448a225664_file-751380c3c945700a14d4c0b3c940618fc1584255.ipynb

pomegranate-vs-hmmlearn.ipynb

pomegranate-vs-hmmlearn

pomegranate / hmmlearn comparison
hmmlearn is a Python module for hidden markov models with a scikit-learn like API. It was originally present in scikit-learn until its removal due to structural learning not meshing well with the API of many other classical machine learning algorithms. Here is a table highlighting some of the similarities and differences between the two packages.


Feature
pomegranate
hmmlearn


Graph Structure




Silent States
✓



Optional Explicit End State
✓



Sparse Implementation
✓



Arbitrary Emissions Allowed on States
✓



Discrete/Gaussian/GMM Emissions
✓
✓


Large Library of Other Emissions
✓



Build Model from Matrices
✓
✓


Build Model Node-by-Node
✓



Serialize to JSON
✓



Serialize using Pickle/Joblib

✓


Algorithms




Priors

✓


Sampling
✓
✓


Log Probability Scoring
✓
✓


Forward-Backward Emissions
✓
✓


Forward-Backward Transitions
✓



Viterbi Decoding
✓
✓


MAP Decoding
✓
✓


Baum-Welch Training
✓
✓


Viterbi Training
✓



Labeled Training
✓



Tied Emissions
✓



Tied Transitions
✓



Emission Inertia
✓



Transition Inertia
✓



Emission Freezing
✓
✓


Transition Freezing
✓
✓


Multi-threaded Training
✓
Coming Soon



Just because the two features are implemented doesn't speak to how fast they are. Below we investigate how fast the two packages are in different settings the two have implemented.
Fully Connected Graphs with Multivariate Gaussian Emissions
Lets look at the sample scoring method, viterbi, and Baum-Welch training for fully connected graphs with multivariate Gaussian emisisons. A fully connected graph is one where all states have connections to all other states. This is a case which pomegranate is expected to do poorly due to its sparse implementation, and hmmlearn should shine due to its vectorized implementations.

pomegranate-vs-hmmlearn

Comments