Models in msmbuilder inherit from base classes in scikit-learn, and follow a similar API. Like sklearn, each type of model is a python class. Hyperparameters are passed in via the __init__ method and set as instance attributes.
from msmbuilder.decomposition import tICA
tica = tICA(gamma=0.05)
tica.fit(...)
# change gamma and refit. the old state will be discarded
tica.gamma = 0.01
tica.fit(...)
The heavy lifting to actually fit the model is done in fit(). In msmbuilder the fit() method always accepts a list of 2-dimensional arrays as input data, where each array represents a single timeseries / trajectory and has a shape of (length_of_trajectory, n_features). Some models can also accept a list of MD trajectories (Trajectory) as opposed to a list of arrays.
dataset = [np.load('traj-1-features.npy'), np.load('traj-2-featues.npy')]
assert dataset[0].ndim == 2 and dataset[1].ndim == 2
clusterer = KCenters(n_clusters=100)
clusterer.fit(dataset)
Note
This is different from sklearn. In sklearn, estimators take a single 2D array as input in fit(). Here we use a list of arrays or trajectories.
Some models like tICA which only require a single pass over the data can also be fit in a potentially more memory efficient way, using the partial_fit method, which can incrementally learn the model one trajectory at a time.
Parameters of the model which are learned or estimated during fit() are always set as instance attributes that are named with a trailing underscore
tica = tICA(gamma=0.05)
tica.fit(...)
# timescales is an estimated quantity, so it ends in an underscore
print(tica.timescales_)
Many models also implement a transform() method, which apply a transformation to a dataset. [TODO: WRITE ME]
The models in msmbuilder are designed to work together as part of a sklearn.pipeline.Pipeline
from msmbuilder.cluster import KMeans
from msmbuilder.msm import MarkovStateModel
from sklearn.pipeline import Pipeline
pipeline = Pipeline([
('cluster', KMeans(n_clusters=100)),
('msm', MarkovStateModel())
])
pipeline.fit(dataset)