def set_model_params(self, models=None, f_therm=None, pi=None, f=None, label='ground state'):
# don't normalize f, because in a multiensemble the relative energy levels matter
_StationaryModel.set_model_params(self, pi=pi, f=f, normalize_f=False)
# check and set other parameters
_types.assert_array(f_therm, ndim=1, kind='numeric')
f_therm = _np.array(f_therm, dtype=float)
for m in models:
assert issubclass(m.__class__, _Model)
self.update_model_params(models=models, f_therm=f_therm)
def _estimate(self, trajs):
"""
Parameters
----------
trajs : ndarray(T, 2) or list of ndarray(T_i, 2)
Thermodynamic trajectories. Each trajectory is a (T_i, 2)-array
with T_i time steps. The first column is the thermodynamic state
index, the second column is the configuration state index.
"""
# format input if needed
if isinstance(trajs, _np.ndarray):
trajs = [trajs]
# validate input
assert _types.is_list(trajs)
for ttraj in trajs:
_types.assert_array(ttraj, ndim=2, kind='numeric')
assert _np.shape(ttraj)[1] >= 2
# harvest state counts
self.state_counts_full = _util.state_counts(
[_np.ascontiguousarray(t[:, :2]).astype(_np.intc) for t in trajs],
nthermo=self.nthermo,
nstates=self.nstates_full)
# active set
self.active_set = _np.where(self.state_counts_full.sum(axis=0) > 0)[0]
self.state_counts = _np.ascontiguousarray(
self.state_counts_full[:, self.active_set].astype(_np.intc))
self.bias_energies = _np.ascontiguousarray(
self.bias_energies_full[:, self.active_set], dtype=_np.float64)
# run estimator
self.therm_energies, self.conf_energies, self.increments, self.loglikelihoods = \
_wham.estimate(
self.state_counts, self.bias_energies,
maxiter=self.maxiter, maxerr=self.maxerr,
therm_energies=self.therm_energies, conf_energies=self.conf_energies,
save_convergence_info=self.save_convergence_info)
# get stationary models
models = [
_StationaryModel(
pi=_np.exp(self.therm_energies[K, _np.newaxis] -
self.bias_energies[K, :] - self.conf_energies),
f=self.bias_energies[K, :] + self.conf_energies,
normalize_energy=False,
label="K=%d" % K) for K in range(self.nthermo)
]
# set model parameters to self
self.set_model_params(models=models,
f_therm=self.therm_energies,
f=self.conf_energies)
# done
return self
def _estimate(self, trajs):
# check input
assert isinstance(trajs, (tuple, list))
assert len(trajs) == 2
ttrajs = trajs[0]
dtrajs = trajs[1]
# validate input
for ttraj, dtraj in zip(ttrajs, dtrajs):
_types.assert_array(ttraj, ndim=1, kind='numeric')
_types.assert_array(dtraj, ndim=1, kind='numeric')
assert _np.shape(ttraj)[0] == _np.shape(dtraj)[0]
# harvest state counts
self.state_counts_full = _util.state_counts(ttrajs,
dtrajs,
nthermo=self.nthermo,
nstates=self.nstates_full)
# active set
self.active_set = _np.where(self.state_counts_full.sum(axis=0) > 0)[0]
self.state_counts = _np.ascontiguousarray(
self.state_counts_full[:, self.active_set].astype(_np.intc))
self.bias_energies = _np.ascontiguousarray(
self.bias_energies_full[:, self.active_set], dtype=_np.float64)
# run estimator
pg = _ProgressReporter()
stage = 'WHAM'
with pg.context(stage=stage):
self.therm_energies, self.conf_energies, self.increments, self.loglikelihoods = \
_wham.estimate(
self.state_counts, self.bias_energies,
maxiter=self.maxiter, maxerr=self.maxerr,
therm_energies=self.therm_energies, conf_energies=self.conf_energies,
save_convergence_info=self.save_convergence_info,
callback=_ConvergenceProgressIndicatorCallBack(
pg, stage, self.maxiter, self.maxerr))
# get stationary models
models = [
_StationaryModel(
pi=_np.exp(self.therm_energies[K, _np.newaxis] -
self.bias_energies[K, :] - self.conf_energies),
f=self.bias_energies[K, :] + self.conf_energies,
normalize_energy=False,
label="K=%d" % K) for K in range(self.nthermo)
]
# set model parameters to self
self.set_model_params(models=models,
f_therm=self.therm_energies,
f=self.conf_energies)
# done
return self
def set_model_params(self, pi=None, f=None, normalize_f=True):
r"""
Parameters
----------
pi : ndarray(n)
Stationary distribution. If not already normalized, pi will be
scaled to fulfill :math:`\sum_i \pi_i = 1`. The free energies f
will be computed from pi via :math:`f_i = - \log(\pi_i)`. Only
if normalize_f is True, a constant will be added to ensure
consistency with :math:`\sum_i \pi_i = 1`.
f : ndarray(n)
Discrete-state free energies. If normalized_f = True, a constant
will be added to normalize the stationary distribution. Otherwise
f is left as given.
normalize_f : bool, default=True
If parametrized by free energy f, normalize them such that
:math:`\sum_i \pi_i = 1`, which is achieved by :math:`\log \sum_i \exp(-f_i) = 0`.
label : str, default='ground state'
Human-readable description for the thermodynamic state of this
model. May contain a temperature description, such as '300 K' or
a description of bias energy such as 'unbiased' or 'Umbrella 1'
"""
# check input
if pi is None and f is None:
raise ValueError('Trying to initialize model without parameters:'
' Both pi (stationary distribution)'
'and f (free energy) are None.'
'At least one of them needs to be set.')
# use f with preference
if f is not None:
_types.assert_array(f, ndim=1, kind='numeric')
f = _np.array(f, dtype=float)
if normalize_f:
f += _logsumexp(
-f
) # normalize on the level on energies to achieve sum_i pi_i = 1
pi = _np.exp(-f)
else: # if f is not given, use pi. pi can't be None at this point
_types.assert_array(pi, ndim=1, kind='numeric')
pi = _np.array(pi, dtype=float)
f = -_np.log(pi)
pi /= pi.sum() # always normalize pi
# set parameters
self.update_model_params(pi=pi, f=f, normalize_energy=normalize_f)
# set derived quantities
self._nstates = len(pi)
def set_model_params(self, pi=None, f=None, normalize_f=None, label=None):
r"""Call to set all basic model parameters.
Parameters
----------
pi : ndarray(n)
Stationary distribution. If not already normalized, pi will be
scaled to fulfill :math:`\sum_i \pi_i = 1`. The free energies f
will then be computed from pi via :math:`f_i = - \log(\pi_i)`.
f : ndarray(n)
Discrete-state free energies. If normalized_f = True, a constant
will be added to normalize the stationary distribution. Otherwise
f is left as given. Then, pi will be computed from f via :math:`\pi_i = \exp(-f_i)`
and, if necessary, scaled to fulfill :math:`\sum_i \pi_i = 1`. If
both (pi and f) are given, f takes precedence over pi.
normalize_energy : bool, default=True
If parametrized by free energy f, normalize them such that
:math:`\sum_i \pi_i = 1`, which is achieved by :math:`\log \sum_i \exp(-f_i) = 0`.
label : str, default=None
Human-readable description for the thermodynamic state of this
model. May contain a temperature description, such as '300 K' or
a description of bias energy such as 'unbiased' or 'Umbrella 1'.
"""
if f is not None:
_types.assert_array(f, ndim=1, kind='numeric')
f = _np.array(f, dtype=float)
if normalize_f:
f += _logsumexp(
-f
) # normalize on the level on energies to achieve sum_i pi_i = 1
pi = _np.exp(-f)
elif pi is not None: # if f is not given, use pi. pi can't be None at this point
_types.assert_array(pi, ndim=1, kind='numeric')
pi = _np.array(pi, dtype=float)
f = -_np.log(pi)
f += _logsumexp(-f) # always shift f when set by pi
else:
raise ValueError(
"Trying to initialize model without parameters: both pi (stationary distribution)" \
" and f (free energy) are None. At least one of them needs to be set.")
# set parameters (None does not overwrite)
self.update_model_params(pi=pi,
f=f,
normalize_energy=normalize_f,
label=label)
def _estimate(self, trajs):
# check input
assert isinstance(trajs, (tuple, list))
assert len(trajs) == 2
ttrajs = trajs[0]
dtrajs = trajs[1]
# validate input
for ttraj, dtraj in zip(ttrajs, dtrajs):
_types.assert_array(ttraj, ndim=1, kind='numeric')
_types.assert_array(dtraj, ndim=1, kind='numeric')
assert _np.shape(ttraj)[0] == _np.shape(dtraj)[0]
# harvest transition counts
self.count_matrices_full = _util.count_matrices(
ttrajs,
dtrajs,
self.lag,
sliding=self.count_mode,
sparse_return=False,
nstates=self.nstates_full)
# harvest state counts (for WHAM)
self.state_counts_full = _util.state_counts(ttrajs,
dtrajs,
nthermo=self.nthermo,
nstates=self.nstates_full)
# restrict to connected set
C_sum = self.count_matrices_full.sum(axis=0)
# TODO: use improved cset
_, cset = _cset.compute_csets_dTRAM(self.connectivity,
self.count_matrices_full)
self.active_set = cset
# correct counts
self.count_matrices = self.count_matrices_full[:, cset[:, _np.newaxis],
cset]
self.count_matrices = _np.require(self.count_matrices,
dtype=_np.intc,
requirements=['C', 'A'])
# correct bias matrix
self.bias_energies = self.bias_energies_full[:, cset]
self.bias_energies = _np.require(self.bias_energies,
dtype=_np.float64,
requirements=['C', 'A'])
# correct state counts
self.state_counts = self.state_counts_full[:, cset]
self.state_counts = _np.require(self.state_counts,
dtype=_np.intc,
requirements=['C', 'A'])
# run initialisation
pg = _ProgressReporter()
if self.init is not None and self.init == 'wham':
stage = 'WHAM init.'
with pg.context(stage=stage):
self.therm_energies, self.conf_energies, _increments, _loglikelihoods = \
_wham.estimate(
self.state_counts, self.bias_energies,
maxiter=self.init_maxiter, maxerr=self.init_maxerr, save_convergence_info=0,
therm_energies=self.therm_energies, conf_energies=self.conf_energies,
callback=_ConvergenceProgressIndicatorCallBack(
pg, stage, self.init_maxiter, self.init_maxerr))
# run estimator
stage = 'DTRAM'
with pg.context(stage=stage):
self.therm_energies, self.conf_energies, self.log_lagrangian_mult, \
self.increments, self.loglikelihoods = _dtram.estimate(
self.count_matrices, self.bias_energies,
maxiter=self.maxiter, maxerr=self.maxerr,
log_lagrangian_mult=self.log_lagrangian_mult,
conf_energies=self.conf_energies,
save_convergence_info=self.save_convergence_info,
callback=_ConvergenceProgressIndicatorCallBack(
pg, stage, self.maxiter, self.maxerr))
# compute models
fmsms = [
_dtram.estimate_transition_matrix(
self.log_lagrangian_mult, self.bias_energies,
self.conf_energies, self.count_matrices,
_np.zeros(shape=self.conf_energies.shape,
dtype=_np.float64), K) for K in range(self.nthermo)
]
active_sets = [
_largest_connected_set(msm, directed=False) for msm in fmsms
]
fmsms = [
_np.ascontiguousarray((msm[lcc, :])[:, lcc])
for msm, lcc in zip(fmsms, active_sets)
]
models = []
for i, (msm, acs) in enumerate(zip(fmsms, active_sets)):
models.append(
_ThermoMSM(
msm,
self.active_set[acs],
self.nstates_full,
pi=_np.exp(self.therm_energies[i] -
self.bias_energies[i, :] - self.conf_energies),
dt_model=self.timestep_traj.get_scaled(self.lag)))
# set model parameters to self
self.set_model_params(models=models,
f_therm=self.therm_energies,
f=self.conf_energies)
# done
return self
def _estimate(self, X):
ttrajs, dtrajs_full, btrajs = X
# shape and type checks
assert len(ttrajs) == len(dtrajs_full) == len(btrajs)
for t in ttrajs:
_types.assert_array(t, ndim=1, kind='i')
for d in dtrajs_full:
_types.assert_array(d, ndim=1, kind='i')
for b in btrajs:
_types.assert_array(b, ndim=2, kind='f')
# find dimensions
self.nstates_full = max(_np.max(d) for d in dtrajs_full) + 1
self.nthermo = max(_np.max(t) for t in ttrajs) + 1
# dimensionality checks
for t, d, b, in zip(ttrajs, dtrajs_full, btrajs):
assert t.shape[0] == d.shape[0] == b.shape[0]
assert b.shape[1] == self.nthermo
# cast types and change axis order if needed
ttrajs = [
_np.require(t, dtype=_np.intc, requirements='C') for t in ttrajs
]
dtrajs_full = [
_np.require(d, dtype=_np.intc, requirements='C')
for d in dtrajs_full
]
btrajs = [
_np.require(b, dtype=_np.float64, requirements='C') for b in btrajs
]
# find state visits
self.state_counts_full = _util.state_counts(ttrajs, dtrajs_full)
self.therm_state_counts_full = self.state_counts_full.sum(axis=1)
self.active_set = _np.sort(
_np.where(self.state_counts_full.sum(axis=0) > 0)[0])
self.state_counts = _np.ascontiguousarray(
self.state_counts_full[:, self.active_set].astype(_np.intc))
if self.direct_space:
mbar = _mbar_direct
else:
mbar = _mbar
pg = _ProgressReporter()
with pg.context():
self.therm_energies, self.unbiased_conf_energies_full, self.biased_conf_energies_full, \
self.increments = mbar.estimate(
self.state_counts_full.sum(axis=1), btrajs, dtrajs_full,
maxiter=self.maxiter, maxerr=self.maxerr,
save_convergence_info=self.save_convergence_info,
callback=_ConvergenceProgressIndicatorCallBack(
pg, 'MBAR', self.maxiter, self.maxerr),
n_conf_states=self.nstates_full)
try:
self.loglikelihoods = _np.nan * self.increments
except TypeError:
self.loglikelihoods = None
# get stationary models
models = [
_StationaryModel(f=self.biased_conf_energies_full[K,
self.active_set],
normalize_energy=False,
label="K=%d" % K) for K in range(self.nthermo)
]
# set model parameters to self
self.set_model_params(
models=models,
f_therm=self.therm_energies,
f=self.unbiased_conf_energies_full[self.active_set])
self.btrajs = btrajs
# done
return self
def _estimate(self, X):
ttrajs, dtrajs_full, btrajs = X
# shape and type checks
assert len(ttrajs) == len(dtrajs_full) == len(btrajs)
for t in ttrajs:
_types.assert_array(t, ndim=1, kind='i')
for d in dtrajs_full:
_types.assert_array(d, ndim=1, kind='i')
for b in btrajs:
_types.assert_array(b, ndim=2, kind='f')
# find dimensions
nstates_full = max(_np.max(d) for d in dtrajs_full) + 1
if self.nstates_full is None:
self.nstates_full = nstates_full
elif self.nstates_full < nstates_full:
raise RuntimeError("Found more states (%d) than specified by nstates_full (%d)" % (
nstates_full, self.nstates_full))
self.nthermo = max(_np.max(t) for t in ttrajs) + 1
# dimensionality checks
for t, d, b, in zip(ttrajs, dtrajs_full, btrajs):
assert t.shape[0] == d.shape[0] == b.shape[0]
assert b.shape[1] == self.nthermo
# cast types and change axis order if needed
ttrajs = [_np.require(t, dtype=_np.intc, requirements='C') for t in ttrajs]
dtrajs_full = [_np.require(d, dtype=_np.intc, requirements='C') for d in dtrajs_full]
btrajs = [_np.require(b, dtype=_np.float64, requirements='C') for b in btrajs]
# if equilibrium information is given, separate the trajectories
if self.equilibrium is not None:
assert len(self.equilibrium) == len(ttrajs)
_ttrajs, _dtrajs_full, _btrajs = ttrajs, dtrajs_full, btrajs
ttrajs = [ttraj for eq, ttraj in zip(self.equilibrium, _ttrajs) if not eq]
dtrajs_full = [dtraj for eq, dtraj in zip(self.equilibrium, _dtrajs_full) if not eq]
self.btrajs = [btraj for eq, btraj in zip(self.equilibrium, _btrajs) if not eq]
equilibrium_ttrajs = [ttraj for eq, ttraj in zip(self.equilibrium, _ttrajs) if eq]
equilibrium_dtrajs_full = [dtraj for eq, dtraj in zip(self.equilibrium, _dtrajs_full) if eq]
self.equilibrium_btrajs = [btraj for eq, btraj in zip(self.equilibrium, _btrajs) if eq]
else: # set dummy values
equilibrium_ttrajs = []
equilibrium_dtrajs_full = []
self.equilibrium_btrajs = []
self.btrajs = btrajs
# find state visits and transition counts
state_counts_full = _util.state_counts(ttrajs, dtrajs_full, nstates=self.nstates_full, nthermo=self.nthermo)
count_matrices_full = _util.count_matrices(ttrajs, dtrajs_full,
self.lag, sliding=self.count_mode, sparse_return=False, nstates=self.nstates_full, nthermo=self.nthermo)
self.therm_state_counts_full = state_counts_full.sum(axis=1)
if self.equilibrium is not None:
self.equilibrium_state_counts_full = _util.state_counts(equilibrium_ttrajs, equilibrium_dtrajs_full,
nstates=self.nstates_full, nthermo=self.nthermo)
else:
self.equilibrium_state_counts_full = _np.zeros((self.nthermo, self.nstates_full), dtype=_np.float64)
pg = _ProgressReporter()
stage = 'cset'
with pg.context(stage=stage):
self.csets, pcset = _cset.compute_csets_TRAM(
self.connectivity, state_counts_full, count_matrices_full,
equilibrium_state_counts=self.equilibrium_state_counts_full,
ttrajs=ttrajs+equilibrium_ttrajs, dtrajs=dtrajs_full+equilibrium_dtrajs_full, bias_trajs=self.btrajs+self.equilibrium_btrajs,
nn=self.nn, factor=self.connectivity_factor,
callback=_IterationProgressIndicatorCallBack(pg, 'finding connected set', stage=stage))
self.active_set = pcset
# check for empty states
for k in range(self.nthermo):
if len(self.csets[k]) == 0:
_warnings.warn(
'Thermodynamic state %d' % k \
+ ' contains no samples after reducing to the connected set.', EmptyState)
# deactivate samples not in the csets, states are *not* relabeled
self.state_counts, self.count_matrices, self.dtrajs, _ = _cset.restrict_to_csets(
self.csets,
state_counts=state_counts_full, count_matrices=count_matrices_full,
ttrajs=ttrajs, dtrajs=dtrajs_full)
if self.equilibrium is not None:
self.equilibrium_state_counts, _, self.equilibrium_dtrajs, _ = _cset.restrict_to_csets(
self.csets,
state_counts=self.equilibrium_state_counts_full, ttrajs=equilibrium_ttrajs, dtrajs=equilibrium_dtrajs_full)
else:
self.equilibrium_state_counts = _np.zeros((self.nthermo, self.nstates_full), dtype=_np.intc) # (remember: no relabeling)
self.equilibrium_dtrajs = []
# self-consistency tests
assert _np.all(self.state_counts >= _np.maximum(self.count_matrices.sum(axis=1), \
self.count_matrices.sum(axis=2)))
assert _np.all(_np.sum(
[_np.bincount(d[d>=0], minlength=self.nstates_full) for d in self.dtrajs],
axis=0) == self.state_counts.sum(axis=0))
assert _np.all(_np.sum(
[_np.bincount(t[d>=0], minlength=self.nthermo) for t, d in zip(ttrajs, self.dtrajs)],
axis=0) == self.state_counts.sum(axis=1))
if self.equilibrium is not None:
assert _np.all(_np.sum(
[_np.bincount(d[d >= 0], minlength=self.nstates_full) for d in self.equilibrium_dtrajs],
axis=0) == self.equilibrium_state_counts.sum(axis=0))
assert _np.all(_np.sum(
[_np.bincount(t[d >= 0], minlength=self.nthermo) for t, d in zip(equilibrium_ttrajs, self.equilibrium_dtrajs)],
axis=0) == self.equilibrium_state_counts.sum(axis=1))
# check for empty states
for k in range(self.state_counts.shape[0]):
if self.count_matrices[k, :, :].sum() == 0 and self.equilibrium_state_counts[k, :].sum()==0:
_warnings.warn(
'Thermodynamic state %d' % k \
+ ' contains no transitions and no equilibrium data after reducing to the connected set.', EmptyState)
if self.init == 'mbar' and self.biased_conf_energies is None:
if self.direct_space:
mbar = _mbar_direct
else:
mbar = _mbar
stage = 'MBAR init.'
with pg.context(stage=stage):
self.mbar_therm_energies, self.mbar_unbiased_conf_energies, \
self.mbar_biased_conf_energies, _ = mbar.estimate(
(state_counts_full.sum(axis=1)+self.equilibrium_state_counts_full.sum(axis=1)).astype(_np.intc),
self.btrajs+self.equilibrium_btrajs, dtrajs_full+equilibrium_dtrajs_full,
maxiter=self.init_maxiter, maxerr=self.init_maxerr,
callback=_ConvergenceProgressIndicatorCallBack(
pg, stage, self.init_maxiter, self.init_maxerr),
n_conf_states=self.nstates_full)
self.biased_conf_energies = self.mbar_biased_conf_energies.copy()
# run estimator
if self.direct_space:
tram = _tram_direct
trammbar = _trammbar_direct
else:
tram = _tram
trammbar = _trammbar
#import warnings
#with warnings.catch_warnings() as cm:
# warnings.filterwarnings('ignore', RuntimeWarning)
stage = 'TRAM'
with pg.context(stage=stage):
if self.equilibrium is None:
self.biased_conf_energies, conf_energies, self.therm_energies, self.log_lagrangian_mult, \
self.increments, self.loglikelihoods = tram.estimate(
self.count_matrices, self.state_counts, self.btrajs, self.dtrajs,
maxiter=self.maxiter, maxerr=self.maxerr,
biased_conf_energies=self.biased_conf_energies,
log_lagrangian_mult=self.log_lagrangian_mult,
save_convergence_info=self.save_convergence_info,
callback=_ConvergenceProgressIndicatorCallBack(
pg, stage, self.maxiter, self.maxerr, subcallback=self.callback),
N_dtram_accelerations=self.N_dtram_accelerations)
else: # use trammbar
self.biased_conf_energies, conf_energies, self.therm_energies, self.log_lagrangian_mult, \
self.increments, self.loglikelihoods = trammbar.estimate(
self.count_matrices, self.state_counts, self.btrajs, self.dtrajs,
equilibrium_therm_state_counts=self.equilibrium_state_counts.sum(axis=1).astype(_np.intc),
equilibrium_bias_energy_sequences=self.equilibrium_btrajs, equilibrium_state_sequences=self.equilibrium_dtrajs,
maxiter=self.maxiter, maxerr=self.maxerr,
save_convergence_info=self.save_convergence_info,
biased_conf_energies=self.biased_conf_energies,
log_lagrangian_mult=self.log_lagrangian_mult,
callback=_ConvergenceProgressIndicatorCallBack(
pg, stage, self.maxiter, self.maxerr, subcallback=self.callback),
N_dtram_accelerations=self.N_dtram_accelerations,
overcounting_factor=self.overcounting_factor)
# compute models
fmsms = [_np.ascontiguousarray((
_tram.estimate_transition_matrix(
self.log_lagrangian_mult, self.biased_conf_energies, self.count_matrices, None,
K)[self.active_set, :])[:, self.active_set]) for K in range(self.nthermo)]
active_sets = [_largest_connected_set(msm, directed=False) for msm in fmsms]
fmsms = [_np.ascontiguousarray(
(msm[lcc, :])[:, lcc]) for msm, lcc in zip(fmsms, active_sets)]
models = []
for i, (msm, acs) in enumerate(zip(fmsms, active_sets)):
pi_acs = _np.exp(self.therm_energies[i] - self.biased_conf_energies[i, :])[self.active_set[acs]]
pi_acs = pi_acs / pi_acs.sum()
models.append(_ThermoMSM(
msm, self.active_set[acs], self.nstates_full, pi=pi_acs,
dt_model=self.timestep_traj.get_scaled(self.lag)))
# set model parameters to self
self.set_model_params(
models=models, f_therm=self.therm_energies, f=conf_energies[self.active_set].copy())
return self
def _estimate(self, X):
ttrajs, dtrajs_full, btrajs = X
# shape and type checks
assert len(ttrajs) == len(dtrajs_full) == len(btrajs)
for t in ttrajs:
_types.assert_array(t, ndim=1, kind='i')
for d in dtrajs_full:
_types.assert_array(d, ndim=1, kind='i')
for b in btrajs:
_types.assert_array(b, ndim=2, kind='f')
# find dimensions
self.nstates_full = max(_np.max(d) for d in dtrajs_full) + 1
self.nthermo = max(_np.max(t) for t in ttrajs) + 1
# dimensionality checks
for t, d, b, in zip(ttrajs, dtrajs_full, btrajs):
assert t.shape[0] == d.shape[0] == b.shape[0]
assert b.shape[1] == self.nthermo
# cast types and change axis order if needed
ttrajs = [
_np.require(t, dtype=_np.intc, requirements='C') for t in ttrajs
]
dtrajs_full = [
_np.require(d, dtype=_np.intc, requirements='C')
for d in dtrajs_full
]
btrajs = [
_np.require(b, dtype=_np.float64, requirements='C') for b in btrajs
]
# find state visits and transition counts
state_counts_full = _util.state_counts(ttrajs, dtrajs_full)
count_matrices_full = _util.count_matrices(ttrajs,
dtrajs_full,
self.lag,
sliding=self.count_mode,
sparse_return=False,
nstates=self.nstates_full)
self.therm_state_counts_full = state_counts_full.sum(axis=1)
self.csets, pcset = _cset.compute_csets_TRAM(
self.connectivity,
state_counts_full,
count_matrices_full,
ttrajs=ttrajs,
dtrajs=dtrajs_full,
bias_trajs=btrajs,
nn=self.nn,
factor=self.connectivity_factor,
callback=_IterationProgressIndicatorCallBack(
self, 'finding connected set', 'cset'))
self.active_set = pcset
# check for empty states
for k in range(self.nthermo):
if len(self.csets[k]) == 0:
_warnings.warn(
'Thermodynamic state %d' % k \
+ ' contains no samples after reducing to the connected set.', EmptyState)
# deactivate samples not in the csets, states are *not* relabeled
self.state_counts, self.count_matrices, self.dtrajs, _ = _cset.restrict_to_csets(
self.csets,
state_counts=state_counts_full,
count_matrices=count_matrices_full,
ttrajs=ttrajs,
dtrajs=dtrajs_full)
# self-consistency tests
assert _np.all(self.state_counts >= _np.maximum(self.count_matrices.sum(axis=1), \
self.count_matrices.sum(axis=2)))
assert _np.all(
_np.sum([
_np.bincount(d[d >= 0], minlength=self.nstates_full)
for d in self.dtrajs
],
axis=0) == self.state_counts.sum(axis=0))
assert _np.all(
_np.sum([
_np.bincount(t[d >= 0], minlength=self.nthermo)
for t, d in zip(ttrajs, self.dtrajs)
],
axis=0) == self.state_counts.sum(axis=1))
# check for empty states
for k in range(self.state_counts.shape[0]):
if self.count_matrices[k, :, :].sum() == 0:
_warnings.warn(
'Thermodynamic state %d' % k \
+ 'contains no transitions after reducing to the connected set.', EmptyState)
if self.init == 'mbar' and self.biased_conf_energies is None:
if self.direct_space:
mbar = _mbar_direct
else:
mbar = _mbar
self.mbar_therm_energies, self.mbar_unbiased_conf_energies, \
self.mbar_biased_conf_energies, _ = mbar.estimate(
state_counts_full.sum(axis=1), btrajs, dtrajs_full,
maxiter=self.init_maxiter, maxerr=self.init_maxerr,
callback=_ConvergenceProgressIndicatorCallBack(
self, 'MBAR init.', self.init_maxiter, self.init_maxerr),
n_conf_states=self.nstates_full)
self._progress_force_finish(stage='MBAR init.',
description='MBAR init.')
self.biased_conf_energies = self.mbar_biased_conf_energies.copy()
# run estimator
if self.direct_space:
tram = _tram_direct
else:
tram = _tram
#import warnings
#with warnings.catch_warnings() as cm:
# warnings.filterwarnings('ignore', RuntimeWarning)
self.biased_conf_energies, conf_energies, self.therm_energies, self.log_lagrangian_mult, \
self.increments, self.loglikelihoods = tram.estimate(
self.count_matrices, self.state_counts, btrajs, self.dtrajs,
maxiter=self.maxiter, maxerr=self.maxerr,
biased_conf_energies=self.biased_conf_energies,
log_lagrangian_mult=self.log_lagrangian_mult,
save_convergence_info=self.save_convergence_info,
callback=_ConvergenceProgressIndicatorCallBack(
self, 'TRAM', self.maxiter, self.maxerr),
N_dtram_accelerations=self.N_dtram_accelerations)
self._progress_force_finish(stage='TRAM', description='TRAM')
self.btrajs = btrajs
# compute models
fmsms = [
_np.ascontiguousarray((_tram.estimate_transition_matrix(
self.log_lagrangian_mult, self.biased_conf_energies,
self.count_matrices, None,
K)[self.active_set, :])[:, self.active_set])
for K in range(self.nthermo)
]
self.model_active_set = [
_largest_connected_set(msm, directed=False) for msm in fmsms
]
fmsms = [
_np.ascontiguousarray((msm[lcc, :])[:, lcc])
for msm, lcc in zip(fmsms, self.model_active_set)
]
models = [
_MSM(msm, dt_model=self.timestep_traj.get_scaled(self.lag))
for msm in fmsms
]
# set model parameters to self
self.set_model_params(models=models,
f_therm=self.therm_energies,
f=conf_energies[self.active_set].copy())
return self
def _estimate(self, trajs):
"""
Parameters
----------
trajs : ndarray(T, 2) or list of ndarray(T_i, 2)
Thermodynamic trajectories. Each trajectory is a (T_i, 2)-array
with T_i time steps. The first column is the thermodynamic state
index, the second column is the configuration state index.
"""
# format input if needed
if isinstance(trajs, _np.ndarray):
trajs = [trajs]
# validate input
assert _types.is_list(trajs)
for ttraj in trajs:
_types.assert_array(ttraj, ndim=2, kind='numeric')
assert _np.shape(ttraj)[1] >= 2
# harvest transition counts
self.count_matrices_full = _util.count_matrices(
[_np.ascontiguousarray(t[:, :2]).astype(_np.intc) for t in trajs],
self.lag,
sliding=self.count_mode,
sparse_return=False,
nstates=self.nstates_full)
# harvest state counts (for WHAM)
self.state_counts_full = _util.state_counts(trajs,
nthermo=self.nthermo,
nstates=self.nstates_full)
# restrict to connected set
C_sum = self.count_matrices_full.sum(axis=0)
# TODO: use improved cset
cset = _largest_connected_set(C_sum, directed=True)
self.active_set = cset
# correct counts
self.count_matrices = self.count_matrices_full[:, cset[:, _np.newaxis],
cset]
self.count_matrices = _np.require(self.count_matrices,
dtype=_np.intc,
requirements=['C', 'A'])
# correct bias matrix
self.bias_energies = self.bias_energies_full[:, cset]
self.bias_energies = _np.require(self.bias_energies,
dtype=_np.float64,
requirements=['C', 'A'])
# correct state counts
self.state_counts = self.state_counts_full[:, cset]
self.state_counts = _np.require(self.state_counts,
dtype=_np.intc,
requirements=['C', 'A'])
# run initialisation
if self.init is not None:
if self.init == 'wham':
self.therm_energies, self.conf_energies, _increments, _loglikelihoods = \
_wham.estimate(
self.state_counts, self.bias_energies,
maxiter=5000, maxerr=1.0E-8, save_convergence_info=0,
therm_energies=self.therm_energies, conf_energies=self.conf_energies)
# run estimator
self.therm_energies, self.conf_energies, self.log_lagrangian_mult, \
self.increments, self.loglikelihoods = _dtram.estimate(
self.count_matrices, self.bias_energies,
maxiter=self.maxiter, maxerr=self.maxerr,
log_lagrangian_mult=self.log_lagrangian_mult,
conf_energies=self.conf_energies,
save_convergence_info=self.save_convergence_info)
# compute models
models = [
_dtram.estimate_transition_matrix(
self.log_lagrangian_mult, self.bias_energies,
self.conf_energies, self.count_matrices,
_np.zeros(shape=self.conf_energies.shape,
dtype=_np.float64), K) for K in range(self.nthermo)
]
self.model_active_set = [
_largest_connected_set(msm, directed=False) for msm in models
]
models = [
_np.ascontiguousarray((msm[lcc, :])[:, lcc])
for msm, lcc in zip(models, self.model_active_set)
]
# set model parameters to self
self.set_model_params(models=[
_MSM(msm, dt_model=self.timestep_traj.get_scaled(self.lag))
for msm in models
],
f_therm=self.therm_energies,
f=self.conf_energies)
# done
return self
def _estimate(self, trajs):
# TODO: fix docstring
"""
Parameters
----------
X : tuple of (ttrajs, dtrajs)
Simulation trajectories. ttrajs contain the indices of the thermodynamic state and
dtrajs contains the indices of the configurational states.
ttrajs : list of numpy.ndarray(X_i, dtype=int)
Every elements is a trajectory (time series). ttrajs[i][t] is the index of the
thermodynamic state visited in trajectory i at time step t.
dtrajs : list of numpy.ndarray(X_i, dtype=int)
dtrajs[i][t] is the index of the configurational state (Markov state) visited in
trajectory i at time step t.
"""
# check input
assert isinstance(trajs, (tuple, list))
assert len(trajs) == 2
ttrajs = trajs[0]
dtrajs = trajs[1]
# validate input
for ttraj, dtraj in zip(ttrajs, dtrajs):
_types.assert_array(ttraj, ndim=1, kind='numeric')
_types.assert_array(dtraj, ndim=1, kind='numeric')
assert _np.shape(ttraj)[0] == _np.shape(dtraj)[0]
# harvest state counts
self.state_counts_full = _util.state_counts(ttrajs,
dtrajs,
nthermo=self.nthermo,
nstates=self.nstates_full)
# active set
self.active_set = _np.where(self.state_counts_full.sum(axis=0) > 0)[0]
self.state_counts = _np.ascontiguousarray(
self.state_counts_full[:, self.active_set].astype(_np.intc))
self.bias_energies = _np.ascontiguousarray(
self.bias_energies_full[:, self.active_set], dtype=_np.float64)
# run estimator
self.therm_energies, self.conf_energies, self.increments, self.loglikelihoods = \
_wham.estimate(
self.state_counts, self.bias_energies,
maxiter=self.maxiter, maxerr=self.maxerr,
therm_energies=self.therm_energies, conf_energies=self.conf_energies,
save_convergence_info=self.save_convergence_info,
callback=_ConvergenceProgressIndicatorCallBack(
self, 'WHAM', self.maxiter, self.maxerr))
self._progress_force_finish(stage='WHAM')
# get stationary models
models = [
_StationaryModel(
pi=_np.exp(self.therm_energies[K, _np.newaxis] -
self.bias_energies[K, :] - self.conf_energies),
f=self.bias_energies[K, :] + self.conf_energies,
normalize_energy=False,
label="K=%d" % K) for K in range(self.nthermo)
]
# set model parameters to self
self.set_model_params(models=models,
f_therm=self.therm_energies,
f=self.conf_energies)
# done
return self
def _estimate(self, trajs):
# check input
assert isinstance(trajs, (tuple, list))
assert len(trajs) == 2
ttrajs = trajs[0]
dtrajs = trajs[1]
# validate input
for ttraj, dtraj in zip(ttrajs, dtrajs):
_types.assert_array(ttraj, ndim=1, kind='numeric')
_types.assert_array(dtraj, ndim=1, kind='numeric')
assert _np.shape(ttraj)[0] == _np.shape(dtraj)[0]
# harvest transition counts
self.count_matrices_full = _util.count_matrices(
ttrajs, dtrajs, self.lag,
sliding=self.count_mode, sparse_return=False, nstates=self.nstates_full)
# harvest state counts (for WHAM)
self.state_counts_full = _util.state_counts(
ttrajs, dtrajs, nthermo=self.nthermo, nstates=self.nstates_full)
# restrict to connected set
C_sum = self.count_matrices_full.sum(axis=0)
# TODO: use improved cset
cset = _largest_connected_set(C_sum, directed=True)
self.active_set = cset
# correct counts
self.count_matrices = self.count_matrices_full[:, cset[:, _np.newaxis], cset]
self.count_matrices = _np.require(
self.count_matrices, dtype=_np.intc ,requirements=['C', 'A'])
# correct bias matrix
self.bias_energies = self.bias_energies_full[:, cset]
self.bias_energies = _np.require(
self.bias_energies, dtype=_np.float64 ,requirements=['C', 'A'])
# correct state counts
self.state_counts = self.state_counts_full[:, cset]
self.state_counts = _np.require(self.state_counts, dtype=_np.intc ,requirements=['C', 'A'])
# run initialisation
if self.init is not None:
if self.init == 'wham':
self.therm_energies, self.conf_energies, _increments, _loglikelihoods = \
_wham.estimate(
self.state_counts, self.bias_energies,
maxiter=self.init_maxiter, maxerr=self.init_maxerr, save_convergence_info=0,
therm_energies=self.therm_energies, conf_energies=self.conf_energies,
callback=_ConvergenceProgressIndicatorCallBack(
self, 'WHAM init.', self.init_maxiter, self.init_maxerr))
self._progress_force_finish(stage='WHAM init.', description='WHAM init.')
# run estimator
self.therm_energies, self.conf_energies, self.log_lagrangian_mult, \
self.increments, self.loglikelihoods = _dtram.estimate(
self.count_matrices, self.bias_energies,
maxiter=self.maxiter, maxerr=self.maxerr,
log_lagrangian_mult=self.log_lagrangian_mult,
conf_energies=self.conf_energies,
save_convergence_info=self.save_convergence_info,
callback=_ConvergenceProgressIndicatorCallBack(
self, 'DTRAM', self.maxiter, self.maxerr))
self._progress_force_finish(stage='DTRAM', description='DTRAM')
# compute models
models = [_dtram.estimate_transition_matrix(
self.log_lagrangian_mult, self.bias_energies, self.conf_energies,
self.count_matrices, _np.zeros(
shape=self.conf_energies.shape, dtype=_np.float64), K) for K in range(self.nthermo)]
self.model_active_set = [_largest_connected_set(msm, directed=False) for msm in models]
models = [_np.ascontiguousarray(
(msm[lcc, :])[:, lcc]) for msm, lcc in zip(models, self.model_active_set)]
# set model parameters to self
self.set_model_params(
models=[_MSM(msm, dt_model=self.timestep_traj.get_scaled(self.lag)) for msm in models],
f_therm=self.therm_energies, f=self.conf_energies)
# done
return self
