def test_load_batch(self):
'''just test ValueError
'''
self.assertRaises(
ValueError, lambda: _load_batch_pmap(n_cores=4,
lines=['autoimage'],
traj=None,
dtype='dict',
root=0,
mode='xyz',
ref=None))
def test_mpi_load_batch():
# create ``comm`` so you can have the info about n_cpus, cpu id
comm = MPI.COMM_WORLD
# you are free to update anything below here
# _load_batch_pmap is temp method name, will be changed in future
traj_name = fn("tz2.nc")
parm_name = fn("tz2.parm7")
# load to TrajectoryIterator
traj = pt.iterload(traj_name, parm_name, frame_slice=(0, 4000))
# make a list of things you want to
lines = ['autoimage', 'distance :3 :10', 'molsurf @CA']
# gather the data to 1st core (rank=0)
#
n_cores = comm.size
data = _load_batch_pmap(n_cores,
lines=lines,
traj=traj,
mode='mpi',
dtype='dict')
if comm.rank != 0:
assert data is None
if comm.rank == 0:
# each core return a tuple (core_id, dict)
# so you need to concat the dict
# use `from pytraj.tools import concat_dict
data_0 = concat_dict(x[0] for x in data)
# assert to serial version (do not need to copy below to your script)
state = pt.load_batch(traj, lines)
state.run()
data = state.data[1:].to_dict()
for key_0, key in zip(sorted(data_0.keys()), sorted(data.keys())):
aa_eq(data_0[key_0], data[key])
def pmap_mpi(func, traj, *args, **kwd):
"""parallel with MPI (mpi4py)
Parameters
----------
func : a function
traj : pytraj.TrajectoryIterator
*args, **kwd: additional arguments
Examples
--------
.. code-block:: bash
$ # create test_radgyr.py file
$ cat > test_radgyr.py <<EOF
import pytraj as pt
from mpi4py import MPI
comm = MPI.COMM_WORLD
traj = pt.iterload('tz2.nc', 'tz2.parm7')
result_arr = pt.pmap_mpi(pt.radgyr, traj, "@CA")
if comm.rank == 0:
# save data to disk to read later by pytraj.read_pickle
# pt.to_pickle(result_arr, 'output.pk')
print(result_arr)
EOF
$ # run in parallel
$ mpirun -n 4 python ./test_radgyr.py
[array([ 8.10916061, 7.7643485 , 8.09693108, ..., 9.70825678,
9.3161563 , 8.86720964]), array([ 8.82037273, 8.89008289, 9.48540176, ..., 9.29585981,
9.53138062, 9.19155977]), array([ 9.13735723, 8.94651001, 8.97810478, ..., 7.68751186,
8.31361647, 7.83763754]), array([ 7.37423766, 7.05637263, 6.52135566, ..., 6.38061648,
6.24139008, 6.48994552])]
"""
from mpi4py import MPI
comm = MPI.COMM_WORLD
n_cores = comm.size
rank = comm.rank
# update reference
if 'ref' in kwd:
kwd['ref'] = get_reference(traj, kwd['ref'])
if not isinstance(func, (list, tuple)):
# split traj to ``n_cores`` chunks, perform calculation
# for rank-th chunk
if 'dtype' not in kwd:
kwd['dtype'] = 'dict'
frame_indices = kwd.pop('frame_indices', None)
if frame_indices is None:
start, stop = split_range(n_cores, 0, traj.n_frames)[rank]
my_iter = traj.iterframe(start=start, stop=stop)
else:
my_indices = np.array_split(frame_indices, n_cores)[rank]
my_iter = traj.iterframe(frame_indices=my_indices)
data = func(my_iter, *args, **kwd)
total = comm.gather(data, root=0)
if rank == 0:
total = concat_dict(x for x in total)
else:
# cpptraj command style
from pytraj.parallel.base import _load_batch_pmap
total = _load_batch_pmap(n_cores=n_cores,
traj=traj,
lines=func,
dtype='dict',
root=0,
mode='mpi',
**kwd)
if rank == 0:
# otherwise, total=None
total = concat_dict((x[1] for x in total))
return total
from pytraj.parallel.base import _load_batch_pmap
root_dir = "../../tests/data/"
traj_name = root_dir + "tz2.nc"
parm_name = root_dir + "tz2.parm7"
# load to TrajectoryIterator
traj = pt.iterload(traj_name, parm_name, frame_slice=(0, 4000))
# make a list of things you want to
lines = ['autoimage', 'distance :3 :10', 'molsurf @CA']
# gather the data to 1st core (rank=0)
#
n_cores = comm.size
data = _load_batch_pmap(n_cores, lines=lines, traj=traj, mode='mpi', dtype='dict')
if comm.rank != 0:
assert data is None
if comm.rank == 0:
# each core return a tuple (core_id, dict)
# so you need to concat the dict
# use `from pytraj.tools import concat_dict
data_0 = concat_dict(x[0] for x in data)
# assert to serial version (do not need to copy below to your script)
state = pt.load_batch(traj, lines)
state.run()
data = state.data[1:].to_dict()
def test_load_batch(self):
'''just test ValueError
'''
self.assertRaises(
ValueError,
lambda: _load_batch_pmap(n_cores=4, lines=['autoimage'], traj=None, dtype='dict', root=0, mode='xyz', ref=None))
def _pmap(func, traj, *args, **kwargs):
'''use python's multiprocessing to accelerate calculation. Limited calculations.
Parameters
----------
func : a pytraj's methods or a list of string or simply as a cpptraj' text
traj : pytraj.TrajectoryIterator
n_cores : int, number of cores to be used, default 2. Specify n_cores=-1 to use all available cores
iter_options : dict, default {}
Specify trajectory iterating option. This will be done before calling ``func``.
frame_indices : {None, array-like}, default None, optional
if provided, pytraj will split this frame_indices into different chunks and let
cpptraj perform calculation for specific indices.
frame_indices must be pickable so is can be sent to different cores.
*args, **kwargs: additional keywords
Returns
-------
out : OrderedDict
Notes
-----
- If you not sure about parallel's results, you should compare the output to serial run.
- This is absolutely experimental. The syntax might be changed in future.
Rule of thumbs: start with small number of frames (saying 10 frames), varying
n_cores=1, 2, 3, 4 to see if the data makes sense or not.
There are two modes in this method, use pytraj's methods (pytraj.rmsd, pytraj.radgyr,
...) or use cpptraj's command text syntax ('autoimage', 'rms', ...)
If using cpptraj syntax::
pytraj only supports limited cpptraj's Actions (not Analysis, checm Amber15 manual
about Action and Analysis), say no to 'matrix', 'atomicfluct', ... or any action
that results output depending on the number of frames.
This method only benifits you if your calculation is quite long (saying few minutes to
few hours). For calculation that takes less than 1 minutes, you won't see the
significant speed up (or even slower) since pytraj need to warm up and need to gather
data when the calculation done.
The parallel cacluation is very simple, trajectory will be split (almost equal) to
different chunk (n_chunks = n_cores), pytraj/cpptraj perform calculation for each
chunk in each core and then send data back to master. Note that we are using Python's
built-in multiprocessing module, so you can use this method interactively in Ipython
and ipython/jupyter notebook. This behavior is different from using MPI, in which you
need to write a script, escaping ipython ession and type something like::
mpirun -n 4 python my_script.py
vs::
In [1]: pt.pmap(pt.radgyr, traj, n_cores=4)
Out[1]:
OrderedDict([('RoG_00000',
array([ 18.91114428, 18.93654996, 18.84969884, 18.90449256,
18.8568644 , 18.88917208, 18.9430491 , 18.88878079,
18.91669565, 18.87069722]))])
This is experimental method, you should expect its syntax, default output will be changed.
When sending Topology to different cores, pytraj will reload Topology from
traj.top.filename, so if you need to update Topology (in the fly), save it to disk and
reload before using ``pytraj.pmap``
Examples
--------
>>> import numpy as np
>>> import pytraj as pt
>>> traj = pt.load_sample_data('tz2')
>>> # use iter_options
>>> iter_options = {'autoimage': True, 'rmsfit': (0, '@CA')}
>>> data = pt.pmap(pt.mean_structure, traj, iter_options=iter_options)
>>> # cpptraj command style
>>> data = pt.pmap(['distance :3 :7', 'vector mask :3 :12'], traj, n_cores=4)
>>> # use reference. Need to explicitly use 'refindex', which is index of reflist
>>> data = pt.pmap(['rms @CA refindex 0'], traj, ref=[traj[3],], n_cores=3)
>>> data
OrderedDict([('RMSD_00001', array([ 2.68820312e-01, 3.11804885e-01, 2.58835452e-01,
9.10475988e-08, 2.93310737e-01, 4.10197322e-01,
3.96226694e-01, 3.66059215e-01, 3.90890362e-01,
4.89180497e-01]))])
>>> # use reference: if not want to use 'refindex', can use 'reference'
>>> # the advantage is you can not specify a list of reference
>>> data = pt.pmap(['rms @CA reference'], traj, ref=[traj[3],], n_cores=3)
>>> data
OrderedDict([('RMSD_00001', array([ 2.68820312e-01, 3.11804885e-01, 2.58835452e-01,
9.10475988e-08, 2.93310737e-01, 4.10197322e-01,
3.96226694e-01, 3.66059215e-01, 3.90890362e-01,
4.89180497e-01]))])
>>> # use different references. Need to explicitly use 'refindex', which is index of reflist
>>> # create a list of references
>>> reflist = traj[3], traj[4]
>>> # make sure to specify `refindex`
>>> # `refindex 0` is equal to `reflist[0]`
>>> # `refindex 1` is equal to `reflist[1]`
>>> data = pt.pmap(['rms @CA refindex 0', 'rms !@H= refindex 1'], traj, ref=reflist, n_cores=2)
>>> # convert to ndarray
>>> data_arr = pt.tools.dict_to_ndarray(data)
>>> # perform parallel calculation with given frame_indices
>>> traj = pt.datafiles.load_tz2()
>>> data = pt.pmap(pt.radgyr, traj, '@CA', frame_indices=range(10, 50), n_cores=4)
>>> # serial version
>>> data = pt.radgyr(traj, '@CA', frame_indices=range(10, 50))
See also
--------
pytraj.pmap_mpi
'''
from multiprocessing import Pool
from pytraj import TrajectoryIterator
n_cores = kwargs.pop('n_cores') if 'n_cores' in kwargs else 2
iter_options = kwargs.pop(
'iter_options') if 'iter_options' in kwargs else {}
apply = kwargs.pop('apply') if 'apply' in kwargs else None
progress = kwargs.pop('progress') if 'progress' in kwargs else None
progress_params = kwargs.pop(
'progress_params') if 'progress_params' in kwargs else dict()
if n_cores <= 0:
# use all available cores
n_cores = cpu_count()
# update reference
if 'ref' in kwargs:
kwargs['ref'] = get_reference(traj, kwargs['ref'])
if isinstance(func, (list, tuple, string_types)):
# assume using _load_batch_pmap
from pytraj.parallel.base import _load_batch_pmap
#check_valid_command(func)
data = _load_batch_pmap(n_cores=n_cores,
traj=traj,
lines=func,
dtype='dict',
root=0,
mode='multiprocessing',
**kwargs)
data = concat_dict((x[0] for x in data))
return data
else:
if not callable(func):
raise ValueError('must callable argument')
# pytraj's method
if not hasattr(func, '_is_parallelizable'):
raise ValueError("this method does not support parallel")
elif not func._is_parallelizable:
raise ValueError("this method does not support parallel")
else:
if hasattr(
func, '_openmp_capability'
) and func._openmp_capability and 'OPENMP' in compiled_info():
raise RuntimeError(
"this method supports both openmp and pmap, but your cpptraj "
"version was installed with openmp. Should not use both openmp and pmap at the "
"same time. In this case, do not use pmap since openmp is more efficient"
)
if not isinstance(traj, TrajectoryIterator):
raise ValueError('only support TrajectoryIterator')
if 'dtype' not in kwargs and func not in [
mean_structure,
matrix.dist,
matrix.idea,
ired_vector_and_matrix,
rotation_matrix,
volmap,
]:
kwargs['dtype'] = 'dict'
# keyword
if func is volmap:
assert kwargs.get('size') is not None, 'must provide "size" value'
p = Pool(n_cores)
pfuncs = partial(worker_by_func,
n_cores=n_cores,
func=func,
traj=traj,
args=args,
kwargs=kwargs,
iter_options=iter_options,
apply=apply,
progress=progress,
progress_params=progress_params)
data = p.map(pfuncs, [rank for rank in range(n_cores)])
p.close()
dataset_processor = PmapDataset(data,
func=func,
kwargs=kwargs,
traj=traj)
return dataset_processor.process()
def _pmap(func, traj, *args, **kwd):
'''use python's multiprocessing to accelerate calculation. Limited calculations.
Parameters
----------
func : a pytraj's methods or a list of string or simply as a cpptraj' text
traj : pytraj.TrajectoryIterator
n_cores : int, number of cores to be used, default 2. Specify n_cores=-1 to use all available cores
iter_options : dict, default {}
Specify trajectory iterating option. This will be done before calling ``func``.
frame_indices : {None, array-like}, default None, optional
if provided, pytraj will split this frame_indices into different chunks and let
cpptraj perform calculation for specific indices.
frame_indices must be pickable so is can be sent to different cores.
*args, **kwd: additional keywords
Returns
-------
out : OrderedDict
Notes
-----
- If you not sure about parallel's results, you should compare the output to serial run.
- This is absolutely experimental. The syntax might be changed in future.
Rule of thumbs: start with small number of frames (saying 10 frames), varying
n_cores=1, 2, 3, 4 to see if the data makes sense or not.
There are two modes in this method, use pytraj's methods (pytraj.rmsd, pytraj.radgyr,
...) or use cpptraj's command text syntax ('autoimage', 'rms', ...)
If using cpptraj syntax::
pytraj only supports limited cpptraj's Actions (not Analysis, checm Amber15 manual
about Action and Analysis), say no to 'matrix', 'atomicfluct', ... or any action
that results output depending on the number of frames.
This method only benifits you if your calculation is quite long (saying few minutes to
few hours). For calculation that takes less than 1 minutes, you won't see the
significant speed up (or even slower) since pytraj need to warm up and need to gather
data when the calculation done.
The parallel cacluation is very simple, trajectory will be split (almost equal) to
different chunk (n_chunks = n_cores), pytraj/cpptraj perform calculation for each
chunk in each core and then send data back to master. Note that we are using Python's
built-in multiprocessing module, so you can use this method interactively in Ipython
and ipython/jupyter notebook. This behavior is different from using MPI, in which you
need to write a script, escaping ipython ession and type something like::
mpirun -n 4 python my_script.py
vs::
In [1]: pt.pmap(pt.radgyr, traj, n_cores=4)
Out[1]:
OrderedDict([('RoG_00000',
array([ 18.91114428, 18.93654996, 18.84969884, 18.90449256,
18.8568644 , 18.88917208, 18.9430491 , 18.88878079,
18.91669565, 18.87069722]))])
This is experimental method, you should expect its syntax, default output will be changed.
When sending Topology to different cores, pytraj will reload Topology from
traj.top.filename, so if you need to update Topology (in the fly), save it to disk and
reload before using ``pytraj.pmap``
Examples
--------
>>> import numpy as np
>>> import pytraj as pt
>>> traj = pt.load_sample_data('tz2')
>>> # use iter_options
>>> iter_options = {'autoimage': True, 'rmsfit': (0, '@CA')}
>>> data = pt.pmap(pt.mean_structure, traj, iter_options=iter_options)
>>> # cpptraj command style
>>> data = pt.pmap(['distance :3 :7', 'vector mask :3 :12'], traj, n_cores=4)
>>> # use reference. Need to explicitly use 'refindex', which is index of reflist
>>> data = pt.pmap(['rms @CA refindex 0'], traj, ref=[traj[3],], n_cores=3)
>>> data
OrderedDict([('RMSD_00001', array([ 2.68820312e-01, 3.11804885e-01, 2.58835452e-01,
9.10475988e-08, 2.93310737e-01, 4.10197322e-01,
3.96226694e-01, 3.66059215e-01, 3.90890362e-01,
4.89180497e-01]))])
>>> # use reference: if not want to use 'refindex', can use 'reference'
>>> # the advantage is you can not specify a list of reference
>>> data = pt.pmap(['rms @CA reference'], traj, ref=[traj[3],], n_cores=3)
>>> data
OrderedDict([('RMSD_00001', array([ 2.68820312e-01, 3.11804885e-01, 2.58835452e-01,
9.10475988e-08, 2.93310737e-01, 4.10197322e-01,
3.96226694e-01, 3.66059215e-01, 3.90890362e-01,
4.89180497e-01]))])
>>> # use different references. Need to explicitly use 'refindex', which is index of reflist
>>> # create a list of references
>>> reflist = traj[3], traj[4]
>>> # make sure to specify `refindex`
>>> # `refindex 0` is equal to `reflist[0]`
>>> # `refindex 1` is equal to `reflist[1]`
>>> data = pt.pmap(['rms @CA refindex 0', 'rms !@H= refindex 1'], traj, ref=reflist, n_cores=2)
>>> # convert to ndarray
>>> data_arr = pt.tools.dict_to_ndarray(data)
>>> # perform parallel calculation with given frame_indices
>>> traj = pt.datafiles.load_tz2()
>>> data = pt.pmap(pt.radgyr, traj, '@CA', frame_indices=range(10, 50), n_cores=4)
>>> # serial version
>>> data = pt.radgyr(traj, '@CA', frame_indices=range(10, 50))
See also
--------
pytraj.pmap_mpi
'''
from multiprocessing import Pool
from pytraj import TrajectoryIterator
n_cores = kwd.pop('n_cores') if 'n_cores' in kwd else 2
iter_options = kwd.pop('iter_options') if 'iter_options' in kwd else {}
apply = kwd.pop('apply') if 'apply' in kwd else None
if n_cores <= 0:
# use all available cores
n_cores = cpu_count()
# update reference
if 'ref' in kwd:
kwd['ref'] = get_reference(traj, kwd['ref'])
if isinstance(func, (list, tuple, string_types)):
# assume using _load_batch_pmap
from pytraj.parallel.base import _load_batch_pmap
#check_valid_command(func)
data = _load_batch_pmap(n_cores=n_cores,
traj=traj,
lines=func,
dtype='dict',
root=0,
mode='multiprocessing',
**kwd)
data = concat_dict((x[1] for x in data))
return data
else:
if not callable(func):
raise ValueError('must callable argument')
# pytraj's method
if not hasattr(func, '_is_parallelizable'):
raise ValueError("this method does not support parallel")
elif not func._is_parallelizable:
raise ValueError("this method does not support parallel")
else:
if hasattr(
func,
'_openmp_capability') and func._openmp_capability and 'OPENMP' in compiled_info(
):
raise RuntimeError(
"this method supports both openmp and pmap, but your cpptraj "
"version was installed with openmp. Should not use both openmp and pmap at the "
"same time. In this case, do not use pmap since openmp is more efficient")
if not isinstance(traj, TrajectoryIterator):
raise ValueError('only support TrajectoryIterator')
if 'dtype' not in kwd and func not in [
mean_structure, matrix.dist, matrix.idea,
ired_vector_and_matrix, rotation_matrix,
volmap,
]:
kwd['dtype'] = 'dict'
# keyword
if func is volmap:
assert kwd.get('size') is not None, 'must provide "size" value'
p = Pool(n_cores)
pfuncs = partial(worker_byfunc,
n_cores=n_cores,
func=func,
traj=traj,
args=args,
kwd=kwd,
iter_options=iter_options,
apply=apply)
data = p.map(pfuncs, [rank for rank in range(n_cores)])
p.close()
if func in [matrix.dist, matrix.idea, volmap]:
mat = np.sum((val[1] * val[2] for val in data)) / traj.n_frames
return mat
elif func in [ired_vector_and_matrix, ]:
# data is a list of (rank, (vectors, matrix), n_frames)
mat = np.sum((val[1][1] * val[2] for val in data)) / traj.n_frames
vecs = np.column_stack(val[1][0] for val in data)
return (vecs, mat)
elif func in [rotation_matrix, ]:
if 'with_rmsd' in kwd.keys() and kwd['with_rmsd']:
# data is a list of (rank, (mat, rmsd), n_frames)
mat = np.row_stack(val[1][0] for val in data)
rmsd_ = np.hstack(val[1][1] for val in data)
return OrderedDict(out=(mat, rmsd_))
else:
mat = np.row_stack(val[1] for val in data)
return OrderedDict(mat=mat)
elif func == mean_structure:
xyz = np.sum((x[2] * x[1].xyz for x in data)) / traj.n_frames
frame = Frame(xyz.shape[0])
frame.xyz[:] = xyz
return frame
else:
return concat_dict((x[1] for x in data))
root_dir = "data/"
traj_name = root_dir + "tz2.nc"
parm_name = root_dir + "tz2.parm7"
# load to TrajectoryIterator
traj = pt.iterload(traj_name, parm_name, frame_slice=(0, 4000))
# make a list of things you want to
lines = ['autoimage', 'distance :3 :10', 'molsurf @CA']
# gather the data to 1st core (rank=0)
#
n_cores = comm.size
data = _load_batch_pmap(n_cores,
lines=lines,
traj=traj,
mode='mpi',
dtype='dict')
if comm.rank != 0:
assert data is None
if comm.rank == 0:
from pytraj.tools import concat_dict
# each core return a tuple (core_id, dict)
# so you need to concat the dict
# use `from pytraj.tools import concat_dict
data_0 = concat_dict(x[1] for x in data)
# assert to serial version (do not need to copy below to your script)
state = pt.load_batch(traj, lines)