def run(nstates, nsamples):
# Load force data.
from netCDF4 import Dataset
ncfile = Dataset('fiber3-trace011.nc', 'r')
tau = 0.001 # 1 kHz
obs_label = 'force / pN'
time_units = 's' # seconds
o_t = ncfile.variables['force'] # load trace
# force to make a copy because netCDF appears to cause problems
nsubsample = 50 # subsampling rate
o_t = o_t[::nsubsample]
tau *= nsubsample
# -------------------
O = [o_t] # form list of traces
# Initialize MLHMM.
print "Initializing MLHMM with "+str(nstates)+" states."
estimator = bhmm.MLHMM(O, nstates)
# Plot initial guess.
plots.plot_state_assignments(estimator.hmm, None, O[0], time_units=time_units, obs_label=obs_label, tau=tau,
pdf_filename='fiber3-trace11-guess-stateassignments-nstate'+str(nstates)+'.pdf')
# Fit MLHMM
mle = estimator.fit()
# Plot.
plots.plot_state_assignments(mle, mle.hidden_state_trajectories[0], o_t, time_units=time_units,
obs_label=obs_label, tau=tau,
pdf_filename='fiber3-trace11-mlhmm-stateassignments-nstate'+str(nstates)+'.pdf')
# Initialize BHMM, using MLHMM model as initial model.
print "Initializing BHMM and running with "+str(nsamples)+" samples."
sampler = bhmm.BHMM(O, nstates, initial_model=mle)
# Sample models.
bhmm_models = sampler.sample(nsamples=nsamples, save_hidden_state_trajectory=False)
# Generate a sample saving a hidden state trajectory.
final_models = sampler.sample(nsamples=1, save_hidden_state_trajectory=True)
# Plot.
model = final_models[0]
s_t = model.hidden_state_trajectories[0]
o_t = O[0]
plots.plot_state_assignments(model, s_t, o_t, time_units=time_units, obs_label=obs_label, tau=tau,
pdf_filename='fiber3-trace11-bhmm-stateassignments-nstate'+str(nstates)+'.pdf')
# write latex table with sample statistics
conf = 0.95
sampled_hmm = bhmm.SampledGaussianHMM(mle, bhmm_models)
generate_latex_table(sampled_hmm, conf=conf, dt=tau, time_unit='s',
caption='BHMM model estimates for RNA hairpin data.',
outfile='p5ab-bhmm-statistics-table-nstate'+str(nstates)+'.tex')
def run(nstates, nsamples):
# Create model.
true_model = testsystems.force_spectroscopy_model()
nstates = true_model.nstates
tau = 0.001 # time interval per observation
# Generate synthetic data.
print "Generating synthetic data..."
[O, S] = true_model.generate_synthetic_observation_trajectories(ntrajectories=1, length=50000)
# DEBUG
print "synthetic observation trajectories:"
print O
print "Total state visits, min_state, max_state:"
print testsystems.total_state_visits(nstates, S)
# Generate MLHMM.
print "Generating MLHMM..."
estimator = bhmm.MLHMM(O, nstates)
print "Initial guess:"
print str(estimator.hmm.output_model)
print estimator.hmm.transition_matrix
print estimator.hmm.stationary_distribution
# Plot initial guess.
s_t = None
o_t = O[0]
plots.plot_state_assignments(estimator.hmm, s_t, o_t, time_units='s', obs_label='force / pN', tau=tau,
pdf_filename='synthetic-three-state-model-guess-nstates'+str(nstates)+'.pdf')
print "Fitting HMM..."
mle = estimator.fit()
# Plot.
s_t = mle.hidden_state_trajectories[0]
import numpy as np
o_t = O[0]
plots.plot_state_assignments(mle, s_t, o_t, time_units='s', obs_label='force / pN', tau=tau,
pdf_filename='synthetic-three-state-model-mlhmm-nstates'+str(nstates)+'.pdf')
# Initialize BHMM with MLHMM model.
print "Sampling models from BHMM..."
sampler = bhmm.BHMM(O, nstates, initial_model=mle)
bhmm_models = sampler.sample(nsamples=nsamples, save_hidden_state_trajectory=False)
# Generate a sample saving a hidden state trajectory.
final_models = sampler.sample(nsamples=1, save_hidden_state_trajectory=True)
# Plot final BHMM sample.
model = final_models[0]
s_t = model.hidden_state_trajectories[0]
o_t = O[0]
plots.plot_state_assignments(model, s_t, o_t, time_units='s', obs_label='force / pN', tau=tau,
pdf_filename='synthetic-three-state-model-bhmm-nstates'+str(nstates)+'.pdf')
# write latex table with sample statistics
conf = 0.95
sampled_hmm = bhmm.SampledGaussianHMM(mle, bhmm_models)
generate_latex_table(sampled_hmm, conf=conf, dt=1, time_unit='step',
caption='Bayesian HMM parameter estimates for synthetic three-state model.',
outfile='synthetic-three-state-model-bhmm-statistics.tex')
def analyze_data(O, nstates, nsamples=1000, nobservations=None):
"""
Analyze the data with the specified number of states.
Parameters
----------
O : numpy.float
observation trajectory
nstates : int
Number of states to use for analysis.
nsamples : int, optional, default=1000
Number of iterations to sample from the Bayesian posterior for the BHMM.
nobservations : int, optional, default=None
If specified, number of observations to use from O.
"""
# Time interval.
tau = 0.001 # time interval (s) for plotting
# Truncate O to number of observations.
if nobservations:
print "Using only %d observations" % nobservations
O = [ o_t[0:nobservations] for o_t in O ]
else:
nobservations = len(O[0])
# Generate MLHMM.
print "Generating MLHMM..."
estimator = bhmm.MLHMM(O, nstates)
print "Initial guess:"
print str(estimator.hmm.output_model)
print estimator.hmm.transition_matrix
print estimator.hmm.stationary_distribution
# Plot initial guess.
s_t = None
o_t = O[0]
filename = os.path.join('figures', 'synthetic-three-state-model-guess-nstates%(nstates)d-nobs%(nobservations)d.pdf' % vars())
plots.plot_state_assignments(estimator.hmm, s_t, o_t, time_units='s', obs_label='force / pN', tau=tau, pdf_filename=filename)
print "Fitting HMM..."
mle = estimator.fit()
# Plot.
s_t = mle.hidden_state_trajectories[0]
import numpy as np
o_t = O[0]
filename = os.path.join('figures', 'synthetic-three-state-model-mlhmm-nstates%(nstates)d-nobs%(nobservations)d.pdf' % vars())
plots.plot_state_assignments(mle, s_t, o_t, time_units='s', obs_label='force / pN', tau=tau, pdf_filename=filename)
# Initialize BHMM with MLHMM model.
print "Sampling models from BHMM..."
sampler = bhmm.BHMM(O, nstates, initial_model=mle)
bhmm_models = sampler.sample(nsamples=nsamples, save_hidden_state_trajectory=False)
# Generate a sample saving a hidden state trajectory.
final_models = sampler.sample(nsamples=1, save_hidden_state_trajectory=True)
# Plot final BHMM sample.
model = final_models[0]
s_t = model.hidden_state_trajectories[0]
o_t = O[0]
filename = os.path.join('figures', 'synthetic-three-state-model-bhmm-nstates%(nstates)d-nobs%(nobservations)d.pdf' % vars())
plots.plot_state_assignments(model, s_t, o_t, time_units='s', obs_label='force / pN', tau=tau, pdf_filename=filename)
return [mle, bhmm_models]