def __init__(self, data, lag, units='frames'):
from pyemma.coordinates import tica
# data.dat.tolist() might be better?
self.data = data
if isinstance(data, Metric):
if units != 'frames':
raise RuntimeError(
'Cannot use delayed projection TICA with units other than frames for now. Report this to HTMD issues.'
)
metr = data
from pyemma.coordinates.transform.tica import TICA
self.tic = TICA(lag)
p = ProgressBar(len(metr.simulations))
for proj in _projectionGenerator(metr, _getNcpus()):
for pro in proj:
self.tic.partial_fit(pro[0])
p.progress(len(proj))
p.stop()
else:
lag = unitconvert(units, 'frames', lag, data.fstep)
if lag == 0:
raise RuntimeError(
'Lag time conversion resulted in 0 frames. Please use a larger lag-time for TICA.'
)
self.tic = tica(data.dat.tolist(), lag=lag)
def __init__(self, data, lag, units='frames', dimensions=None):
from pyemma.coordinates.transform.tica import TICA as TICApyemma
self.data = data
self.dimensions = dimensions
if isinstance(data, Metric): # Memory efficient TICA projecting trajectories on the fly
if units != 'frames':
raise RuntimeError('Cannot use delayed projection TICA with units other than frames for now. Report this to HTMD issues.')
self.tic = TICApyemma(lag)
metr = data
p = ProgressBar(len(metr.simulations))
for proj in _projectionGenerator(metr, _getNcpus()):
for pro in proj:
if pro is None:
continue
if self.dimensions is None:
self.tic.partial_fit(pro[0])
else: # Sub-select dimensions for fitting
self.tic.partial_fit(pro[0][:, self.dimensions])
p.progress(len(proj))
p.stop()
else: # In-memory TICA
lag = unitconvert(units, 'frames', lag, data.fstep)
if lag == 0:
raise RuntimeError('Lag time conversion resulted in 0 frames. Please use a larger lag-time for TICA.')
self.tic = TICApyemma(lag)
if self.dimensions is None:
datalist = data.dat.tolist()
else: # Sub-select dimensions for fitting
datalist = [x[:, self.dimensions].copy() for x in data.dat]
self.tic.fit(datalist)
def _calculateDirectedComponent(self, sims, St, N):
metr = Metric(sims, skip=self.skip)
metr.set(self.goalprojection)
clustermeans = np.zeros(len(N))
k = 0
for proj in _projectionGenerator(metr, _getNcpus()):
for pro in proj:
clustermeans[:np.max(St[k]) + 1] += np.bincount(
St[k],
self.goalfunction(pro[0]).flatten())
k += 1
return clustermeans / N
def project(self, ndim=None):
""" Projects the data object given to the constructor onto the top `ndim` TICA dimensions
Parameters
----------
ndim : int
The number of TICA dimensions we want to project the data on. If None is given it will use choose a number
of dimensions to cover 95% of the kinetic variance.
Returns
-------
dataTica : :class:`MetricData <htmd.metricdata.MetricData>` object
A new :class:`MetricData <htmd.metricdata.MetricData>` object containing the TICA projected data
Example
-------
>>> from htmd.projections.tica import TICA
>>> tica = TICA(data,20)
>>> dataTica = tica.project(5)
"""
if ndim is not None:
# self.tic._dim = ndim # Old way of doing it. Deprecated since pyEMMA 2.1
self.tic.set_params(
dim=ndim) # Change to this in 2.1 pyEMMA version
if isinstance(
self.data,
Metric): # Doesn't project on correct number of dimensions
proj = []
refs = []
fstep = None
metr = self.data
p = ProgressBar(len(metr.simulations))
k = -1
droppedsims = []
for projecteddata in _projectionGenerator(metr, _getNcpus()):
for pro in projecteddata:
k += 1
if pro is None:
droppedsims.append(k)
continue
proj.append(self.tic.transform(pro[0]))
refs.append(pro[1])
if fstep is None:
fstep = pro[2]
p.progress(len(projecteddata))
p.stop()
simlist = self.data.simulations
simlist = np.delete(simlist, droppedsims)
ref = np.array(refs, dtype=object)
#fstep = 0
parent = None
else:
proj = self.tic.get_output()
simlist = self.data.simlist
ref = self.data.ref
fstep = self.data.fstep
parent = self.data
if ndim is None:
logger.info(
'Kept {} dimension(s) to cover 95% of kinetic variance.'.
format(self.tic.dimension()))
from htmd.metricdata import MetricData
datatica = MetricData(dat=np.array(proj, dtype=object),
simlist=simlist,
ref=ref,
fstep=fstep,
parent=parent)
from pandas import DataFrame
types = []
indexes = []
description = []
for i in range(ndim):
types += ['tica']
indexes += [-1]
description += ['TICA dimension {}'.format(i + 1)]
datatica.map = DataFrame({
'type': types,
'indexes': indexes,
'description': description
})
return datatica
def project(self, ndim=None):
""" Projects the data object given to the constructor onto the top `ndim` TICA dimensions
Parameters
----------
ndim : int
The number of TICA dimensions we want to project the data on. If None is given it will use choose a number
of dimensions to cover 95% of the kinetic variance.
Returns
-------
dataTica : :class:`MetricData <htmd.metricdata.MetricData>` object
A new :class:`MetricData <htmd.metricdata.MetricData>` object containing the TICA projected data
Example
-------
>>> from htmd.projections.tica import TICA
>>> tica = TICA(data,20)
>>> dataTica = tica.project(5)
"""
from tqdm import tqdm
if ndim is not None:
self.tic.set_params(dim=ndim)
keepdata = []
keepdim = None
keepdimdesc = None
if isinstance(
self.data, Metric
): # Memory efficient TICA projecting trajectories on the fly
proj = []
refs = []
fstep = None
metr = self.data
k = -1
droppedsims = []
pbar = tqdm(total=len(metr.simulations))
for projecteddata in _projectionGenerator(metr, _getNcpus()):
for pro in projecteddata:
k += 1
if pro is None:
droppedsims.append(k)
continue
if self.dimensions is not None:
numDimensions = pro[0].shape[1]
keepdim = np.setdiff1d(range(numDimensions),
self.dimensions)
keepdata.append(pro[0][:, keepdim])
proj.append(
self.tic.transform(
pro[0][:, self.dimensions]).astype(np.float32)
) # Sub-select dimensions for projecting
else:
proj.append(
self.tic.transform(pro[0]).astype(np.float32))
refs.append(pro[1])
if fstep is None:
fstep = pro[2]
pbar.update(len(projecteddata))
pbar.close()
simlist = self.data.simulations
simlist = np.delete(simlist, droppedsims)
ref = np.array(refs, dtype=object)
parent = None
if self.dimensions is not None:
from htmd.projections.metric import _singleMolfile
from htmd.molecule.molecule import Molecule
(single, molfile) = _singleMolfile(metr.simulations)
if single:
keepdimdesc = metr.getMapping(Molecule(molfile))
keepdimdesc = keepdimdesc.iloc[keepdim]
else:
if ndim is not None and self.data.numDimensions < ndim:
raise RuntimeError(
'TICA cannot increase the dimensionality of your data. Your data has {} dimensions and you requested {} TICA dimensions'
.format(self.data.numDimensions, ndim))
if self.dimensions is not None:
keepdim = np.setdiff1d(range(self.data.numDimensions),
self.dimensions)
keepdata = [x[:, keepdim] for x in self.data.dat]
if self.data.description is not None:
keepdimdesc = self.data.description.iloc[keepdim]
proj = self.tic.get_output()
simlist = self.data.simlist
ref = self.data.ref
fstep = self.data.fstep
parent = self.data
# If TICA is done on a subset of dimensions, combine non-projected data with projected data
if self.dimensions is not None:
newproj = []
for k, t in zip(keepdata, proj):
newproj.append(np.hstack((k, t)))
proj = newproj
if ndim is None:
ndim = self.tic.dimension()
logger.info(
'Kept {} dimension(s) to cover 95% of kinetic variance.'.
format(ndim))
from htmd.metricdata import MetricData
datatica = MetricData(dat=np.array(proj),
simlist=simlist,
ref=ref,
fstep=fstep,
parent=parent)
from pandas import DataFrame
# TODO: Make this messy pandas creation cleaner. I'm sure I can append rows to DataFrame
types = []
indexes = []
description = []
for i in range(ndim):
types += ['tica']
indexes += [-1]
description += ['TICA dimension {}'.format(i + 1)]
datatica.description = DataFrame({
'type': types,
'atomIndexes': indexes,
'description': description
})
if self.dimensions is not None and keepdimdesc is not None: # If TICA is done on a subset of dims
datatica.description = keepdimdesc.append(datatica.description,
ignore_index=True)
return datatica
def project(self, ndim=None):
""" Projects the data object given to the constructor onto the top `ndim` TICA dimensions
Parameters
----------
ndim : int
The number of TICA dimensions we want to project the data on. If None is given it will use choose a number
of dimensions to cover 95% of the kinetic variance.
Returns
-------
dataTica : :class:`MetricData <htmd.metricdata.MetricData>` object
A new :class:`MetricData <htmd.metricdata.MetricData>` object containing the TICA projected data
Example
-------
>>> from htmd.projections.tica import TICA
>>> tica = TICA(data,20)
>>> dataTica = tica.project(5)
"""
if ndim is not None:
self.tic.set_params(dim=ndim)
keepdata = []
keepdim = None
keepdimdesc = None
if isinstance(self.data, Metric): # Memory efficient TICA projecting trajectories on the fly
proj = []
refs = []
fstep = None
metr = self.data
p = ProgressBar(len(metr.simulations))
k = -1
droppedsims = []
for projecteddata in _projectionGenerator(metr, _getNcpus()):
for pro in projecteddata:
k += 1
if pro is None:
droppedsims.append(k)
continue
if self.dimensions is not None:
numDimensions = pro[0].shape[1]
keepdim = np.setdiff1d(range(numDimensions), self.dimensions)
keepdata.append(pro[0][:, keepdim])
proj.append(self.tic.transform(pro[0][:, self.dimensions]).astype(np.float32)) # Sub-select dimensions for projecting
else:
proj.append(self.tic.transform(pro[0]).astype(np.float32))
refs.append(pro[1])
if fstep is None:
fstep = pro[2]
p.progress(len(projecteddata))
p.stop()
simlist = self.data.simulations
simlist = np.delete(simlist, droppedsims)
ref = np.array(refs, dtype=object)
parent = None
if self.dimensions is not None:
from htmd.projections.metric import _singleMolfile
from htmd.molecule.molecule import Molecule
(single, molfile) = _singleMolfile(metr.simulations)
if single:
keepdimdesc = metr.getMapping(Molecule(molfile))
keepdimdesc = keepdimdesc.iloc[keepdim]
else:
if ndim is not None and self.data.numDimensions < ndim:
raise RuntimeError('TICA cannot increase the dimensionality of your data. Your data has {} dimensions and you requested {} TICA dimensions'.format(self.data.numDimensions, ndim))
if self.dimensions is not None:
keepdim = np.setdiff1d(range(self.data.numDimensions), self.dimensions)
keepdata = [x[:, keepdim] for x in self.data.dat]
if self.data.description is not None:
keepdimdesc = self.data.description.iloc[keepdim]
proj = self.tic.get_output()
simlist = self.data.simlist
ref = self.data.ref
fstep = self.data.fstep
parent = self.data
# If TICA is done on a subset of dimensions, combine non-projected data with projected data
if self.dimensions is not None:
newproj = []
for k, t in zip(keepdata, proj):
newproj.append(np.hstack((k, t)))
proj = newproj
if ndim is None:
ndim = self.tic.dimension()
logger.info('Kept {} dimension(s) to cover 95% of kinetic variance.'.format(ndim))
from htmd.metricdata import MetricData
datatica = MetricData(dat=np.array(proj), simlist=simlist, ref=ref, fstep=fstep, parent=parent)
from pandas import DataFrame
# TODO: Make this messy pandas creation cleaner. I'm sure I can append rows to DataFrame
types = []
indexes = []
description = []
for i in range(ndim):
types += ['tica']
indexes += [-1]
description += ['TICA dimension {}'.format(i+1)]
datatica.description = DataFrame({'type': types, 'atomIndexes': indexes, 'description': description})
if self.dimensions is not None and keepdimdesc is not None: # If TICA is done on a subset of dims
datatica.description = keepdimdesc.append(datatica.description, ignore_index=True)
return datatica
