def create_learning_fns(hidden_dim):
"""Util function to create learning fns to get learned eigenvalues."""
learning_fns = collections.OrderedDict()
# pylint: disable=g-long-lambda
learning_fns['AR'] = lambda o, i: get_eig_from_arparams(
arma.fit_ar(o, i, hidden_dim))
# The ARMA_OLS method gives very big errors, hence not included.
# learning_fns['ARMA_OLS'] = lambda o, i: get_eig_from_arparams(
# arma.fit_arma_iter(o, i, hidden_dim))
# learning_fns['ARMA_RLS_0.1'] = lambda o, i: get_eig_from_arparams(
# arma.fit_arma_iter(o, i, hidden_dim, l2_reg=0.1))
learning_fns['ARMA_RLS'] = lambda o, i: get_eig_from_arparams(
arma.fit_arma_iter(o, i, hidden_dim, l2_reg=0.01))
# learning_fns['ARMA_RLS_0.005'] = lambda o, i: get_eig_from_arparams(
# arma.fit_arma_iter(o, i, hidden_dim, l2_reg=0.005))
# learning_fns['ARMA_RLS_0.001'] = lambda o, i: get_eig_from_arparams(
# arma.fit_arma_iter(o, i, hidden_dim, l2_reg=0.001))
if FLAGS.include_slow_methods:
learning_fns['ARMA_MLE'] = lambda o, i: get_eig_from_arparams(
arma.fit_arma_mle(o, i, hidden_dim))
# The LDS_MLE method fails convergence too often, hence not included.
# learning_fns['LDS_MLE'] = lambda o, i: lds.fit_lds_mle(
# o, i, hidden_dim)
learning_fns['LDS_GIBBS'] = lambda o, i: lds.fit_lds_gibbs(
o,
i,
hidden_dim,
num_update_samples=FLAGS.LDS_GIBBS_num_update_samples)
return learning_fns
def create_model_fns(hdim):
"""Util function to create model fns to fit model params to sequences.
Args:
hdim: Guessed hidden dimension for model fitting.
Returns:
A dictionary mapping method names to model_fns. Each model_fn
takes output seq and input seq, and returns fitted model params.
"""
model_fns = collections.OrderedDict()
# Using raw outputs.
# model_fns['raw_output'] = lambda s: _replace_nan_with_0(s.outputs)
# pylint: disable=g-long-lambda
# Pure AR.
model_fns['AR'] = lambda s: arma.fit_ar(s.outputs, None, hdim)
# Iterated regression without regularization and with regularization.
model_fns['ARMA_OLS'] = lambda s: arma.fit_arma_iter(s.outputs, None, hdim)
model_fns['ARMA'] = lambda s: arma.fit_arma_iter(
s.outputs, None, hdim, l2_reg=0.01)
if FLAGS.include_slow_methods:
model_fns['LDS'] = lambda s: lds.fit_lds_gibbs(
s.outputs,
None,
hdim,
num_update_samples=FLAGS.LDS_GIBBS_num_update_samples)
model_fns['ARMA_MLE'] = lambda s: arma.fit_arma_mle(
s.outputs, None, hdim)
if FLAGS.include_LDS_MLE:
model_fns['LDS_MLE'] = lambda s: lds.fit_lds_mle(s.outputs, None, hdim)
return model_fns
def create_model_fns(hdim):
"""Util function to create model fns to fit model params to sequences.
Args:
hdim: Guessed hidden dimension for model fitting.
Returns:
A dictionary mapping method names to model_fns. Each model_fn
takes output seq and input seq, and returns fitted model params.
"""
model_fns = collections.OrderedDict()
# Using raw outputs.
# model_fns['raw_output'] = lambda o, i: o
# pylint: disable=g-long-lambda
# Pure AR.
model_fns['AR'] = lambda o, i: arma.fit_ar(o, i, hdim)
# Iterated regression without regularization and with regularization.
model_fns['ARMA_OLS'] = lambda o, i: arma.fit_arma_iter(o, i, hdim)
model_fns['ARMA_RLS'] = lambda o, i: arma.fit_arma_iter(
o, i, hdim, l2_reg=0.01)
# Fit AR model and cluster based on AR param roots.
# model_fns['AR_roots'] = lambda o, i: arma.get_eig_from_arparams(
# arma.fit_ar(o, i, hdim))
# Fit ARMA model and cluster based on AR param roots.
# model_fns['ARMA_OLS_roots'] = lambda o, i: arma.get_eig_from_arparams(
# arma.fit_arma_iter(o, i, hdim))
# model_fns['ARMA_RLS_roots_0.01'] = lambda o, i: arma.get_eig_from_arparams(
# arma.fit_arma_iter(o, i, hdim, l2_reg=0.01))
if FLAGS.include_LDS_GIBBS:
model_fns['LDS_GIBBS'] = lambda o, i: lds.fit_lds_gibbs(
o, i, hdim, num_update_samples=FLAGS.LDS_GIBBS_num_update_samples)
if FLAGS.include_ARMA_MLE:
model_fns['ARMA_MLE'] = lambda o, i: arma.fit_arma_mle(o, i, hdim)
if FLAGS.include_LDS_MLE:
model_fns['LDS_MLE'] = lambda o, i: lds.fit_lds_mle(o, i, hdim)
return model_fns
