Source code for computation_cache

from __future__ import print_function, division
from abc import abstractmethod, abstractproperty, ABCMeta
from collections import Iterable, defaultdict
from inspect import getargspec
import inspect
import tempfile
from itertools import product
from os import makedirs, mkdir, remove, unlink
import os
from os.path import exists, isdir, split as path_split, sep as path_sep, join as path_join
import re
import shelve
import shutil
import tempfile
import traceback
from warnings import warn
import numpy as np
from numpy.lib.format import read_array_header_1_0, open_memmap, read_magic
import sys
from grendel.chemistry import Molecule
from grendel.util.containers import AliasedDict
from grendel.util.decorators import typecheck, LazyProperty
from functools import partial
from hashlib import md5
import hashlib
import psi4
import atexit
import resource
# Create a trivial molecule so psi4 doesn't freak out about not having a Molecule
#psi4.geometry("H 0 0 0\nH 0 0 1")

RegexType = type(re.compile(''))

#TODO test suite
#TODO move to grendel
#TODO function to find analogous/duplicate data and remove the extra files

xproduct = lambda *args: product(*[xrange(i) for i in args])
lambda_type = type(lambda x: x)

[docs]class CacheInconsistancyWarning(Warning): pass
[docs]class FileMissingWarning(CacheInconsistancyWarning): pass
[docs]class ShapeMismatchWarning(Warning): pass
[docs]class FileOverwriteWarning(Warning): pass
enabled_warnings = AliasedDict({ (ShapeMismatchWarning, "ShapeMismatchWarning") : True, (FileOverwriteWarning, "FileOverwriteWarning") : True, (FileMissingWarning, "FileMissingWarning") : True }) fail_warnings = AliasedDict({ (ShapeMismatchWarning, "ShapeMismatchWarning") : False, (FileOverwriteWarning, "FileOverwriteWarning") : False, (FileMissingWarning, "FileMissingWarning") : False }) #================================================================================# #region | Warning handling | #TODO move this to grendel, generalize
[docs]def enable_warning(w): enabled_warnings[w] = True
[docs]def disable_warning(w): enabled_warnings[w] = False
[docs]def enable_all_warnings(): for w in enabled_warnings: enabled_warnings[w] = True
[docs]def disable_all_warnings(): for w in enabled_warnings: enabled_warnings[w] = False
[docs]def raise_on_warning(w): fail_warnings[w] = True
[docs]def raise_warning(msg, warning_type): if enabled_warnings[warning_type]: warn(msg, warning_type) if fail_warnings[warning_type]: raise RuntimeError("Critical warning raised.") #endregion #================================================================================#
[docs]class UninitializedDependency(object): def __init__(self, init_func, owner, depends_on=None, requires_optional_arguments=None ): self.init_func = init_func self.owner = owner self.initialized = False if depends_on is None: depends_on = [] self.depends_on = depends_on if requires_optional_arguments is None: requires_optional_arguments = [] self.requires_optional_arguments = requires_optional_arguments for req in self.requires_optional_arguments: if not ComputationCache.optional_attributes[req] == ComputationCache.NoDefaultValue: raise ValueError("Can't require that a user-specified value be given for an optional" " attribute that has a default value.")
[docs] def initialize(self, memo=None): if self.initialized: # Should be unreachable, unless someone is messing with things they shouldn't raise ValueError("Call to used UninitializedDependency initialize function") #----------------------------------------# # Check for circular dependencies if memo is None: memo = [self] else: if self in memo: raise ValueError("Circular Dependency") memo.append(self) #----------------------------------------# # Initialize dependencies for dep in self.depends_on: if not hasattr(self.owner, dep): raise AttributeError("Unknown dependent attribute: {0}".format(dep)) attr = getattr(self.owner, dep) if isinstance(attr, UninitializedDependency): setattr(self.owner, dep, attr.initialize(memo)) #----------------------------------------# # Check to make sure all of the necessary # optional arguments (if any) have been # specified by the user in the cached # computation self.owner for req in self.requires_optional_arguments: if self.owner.optional_arguments[req] == ComputationCache.NoDefaultValue: # Try to get the name of the attribute we're looking for, for error purposes my_stack = inspect.stack() dep_name = None # First, see if we called from another uninitialized dependency if my_stack[1][1] == __file__ and my_stack[1][3] == "initialize": caller_locals = inspect.getargvalues(my_stack[1][0]).locals if 'dep' in caller_locals: dep_name = caller_locals['dep'] # Or check to see if we were called from get_lazy_attribute elif my_stack[1][1] == __file__ and my_stack[1][3] == "get_lazy_attribute": argvals = inspect.getargvalues(my_stack[1][0]) argname = argvals.args[1] caller_locals = argvals.locals if argname in caller_locals: dep_name = caller_locals[argname] # Then raise the (hopefully helpful) error raise ValueError("Initialization of attribute '{0}' requires optional CachedComputation" " argument '{1}' to have a user-defined value".format( dep_name if dep_name is not None else "<unknown attribute>", req )) #----------------------------------------# rv = self.init_func(self.owner) self.initialized = True return rv
def __getstate__(self): raise TypeError("UninitializedDependency instances are not pickleable") #================================================================================# #region | DatumGetter subclasses |
[docs]class DatumGetter(object): __metaclass__ = ABCMeta @property def needs(self): if hasattr(self, "_needs"): return self._needs else: return getargspec(self.__call__).args[2:] @needs.setter
[docs] def needs(self, value): self._needs = value
@abstractmethod def __call__(self, *args, **kwargs): pass def __getstate__(self): raise TypeError("Instances of DatumGetter and its subclasses should never be pickled.")
[docs]class SimpleMethodCallGetter(DatumGetter): def __init__(self, obj, method_name, *args): self._needs = [obj] self.method_name = method_name self.method_args = args def __call__(self, **kwargs): if len(kwargs) > 1: raise TypeError("SimpleMethodCallGetter requires exactly one argument") obj = kwargs.values()[0] return getattr(obj, self.method_name)(*self.method_args)
[docs]class DimensionListGetter(SimpleMethodCallGetter): def __call__(self, **kwargs): if len(kwargs) > 1: raise TypeError("SimpleMethodCallGetter requires exactly one argument") obj = kwargs.values()[0] dimrv = getattr(obj, self.method_name)(*self.method_args) size = dimrv.n() rv = [] for i in range(size): rv.append(dimrv[i]) return rv
[docs]class MemmapArrayGetter(DatumGetter): @abstractmethod
[docs] def get_shape(self, *args): pass
[docs]class TEIGetter(MemmapArrayGetter): """ Abstract parent class of AOTEIGetter and MOTEIGetter """ #################### # Class Attributes # #################### VALID_KERNELS = ( "eri", "erf_eri", "f12", "f12_squared", "f12g12", "f12_double_commutator" ) ################## # Initialization # ################## def __init__(self, kernel_name, physicist_notation=False): self.kernel_name = kernel_name if self.kernel_name not in TEIGetter.VALID_KERNELS: raise NameError("Don't know how to get two electron integrals for kernel '{0}'".format( self.kernel_name )) if self.kernel_name == "eri": # Make a special needs that doesn't include the correlation factor self.needs = [n for n in self.needs if n != "correlation_factor"] self.physicist_notation = physicist_notation ################### # Private Methods # ################### def _copy_in_matrix(self, out, psimat, physicist_notation=False): # TODO copy memory instead, if possible # TODO symmetry nbf1, nbf2, nbf3, nbf4 = out.value.shape ary = np.array(psimat) out.value[...] = ary.reshape(out.shape) if physicist_notation: out.value[...] = out.value.transpose([0,2,1,3]) #if physicist_notation: # for p, q, r, s in xproduct(*out.shape): # out.value[p, r, q, s] = psimat[0, p*nbf2 + q, r*nbf4 + s] #else: # for p, q, r, s in xproduct(*out.shape): # out.value[p, q, r, s] = psimat[0, p*nbf2 + q, r*nbf4 + s] def _get_tei(self, out, func_name, mints, correlation_factor, physicist_notation, C1=None, C2=None, C3=None, C4=None): getter_args = [] if C1 is not None: if C2 is None: C2 = C1 if C4 is None: if C3 is not None: C4 = C3 else: C4 = C1 if C3 is None: C3 = C1 getter_args.extend([C1, C2, C3, C4]) mints_getter = getattr(mints, func_name) if "f12" in func_name: getter_args.insert(0, correlation_factor) elif "erf" in func_name: raise NotImplementedError() psimat = mints_getter(*getter_args) self._copy_in_matrix(out, psimat, physicist_notation) return out
[docs]class AOTEIGetter(TEIGetter):
[docs] def get_shape(self, basis): nbf = basis.nbf() return (nbf,)*4
def __call__(self, out, mints, correlation_factor=None): self._get_tei(out, "ao_" + self.kernel_name, mints, correlation_factor, self.physicist_notation )
[docs]class MOTEIGetter(TEIGetter): # TODO noncanonical transformation
[docs] def get_shape(self, reference_wavefunction): nmo = reference_wavefunction.nmo() return (nmo,)*4
def __call__(self, out, mints, mo_coefficients, correlation_factor=None): self._get_tei(out, "mo_" + self.kernel_name, mints, correlation_factor, self.physicist_notation, mo_coefficients )
[docs]class MatrixGetter(MemmapArrayGetter): def _copy_in_matrix(self, out, psimat): # TODO Symmetry #out.value[...] = np.array(psimat) for p, q in xproduct(*out.shape): out.value[p, q] = psimat[0, p, q]
[docs]class MOCoefficientsGetter(MatrixGetter): """ """ #TODO open shells #################### # Class Attributes # #################### ALLOWED_TYPES = [ "canonical", "boys", "pipek_mezey" ] ################## # Initialization # ################## def __init__(self, orbital_type="canonical"): self.orbital_type = orbital_type if self.orbital_type not in MOCoefficientsGetter.ALLOWED_TYPES: raise ValueError("Unknown orbital type '{0}'".format(orbital_type))
[docs] def get_shape(self, reference_wavefunction, basis): return basis.nbf(), reference_wavefunction.nmo()
def _get_coeffs(self, out, basis, C): if self.orbital_type == "canonical": self._copy_in_matrix(out, C) elif self.orbital_type == "boys" or self.orbital_type == "pipek_mezey": localizer = psi4.Localizer.build(self.orbital_type.upper(), basis, C) localizer.localize() if not localizer.converged: raise RuntimeError("Localization didn't converge") self._copy_in_matrix(out, localizer.L) else: raise NotImplementedError() def __call__(self, out, basis, mo_coefficients): return self._get_coeffs(out, basis, mo_coefficients)
[docs]class OccupiedMOCoefficientsGetter(MOCoefficientsGetter):
[docs] def get_shape(self, reference_wavefunction, basis): return basis.nbf(), sum(reference_wavefunction.doccpi())
def __call__(self, out, basis, docc_space): self._get_coeffs(out, basis, docc_space.C())
[docs]class OEIGetter(MemmapArrayGetter): def __init__(self, kernel): self.kernel = kernel if not kernel in self.__class__.VALID_KERNELS: raise NameError("Don't know how to get one electron integrals for kernel '{0}'".format( self.kernel )) #DEPRECATED: Chunks data incorrectly
[docs]class AOMultipoleGetter(MemmapArrayGetter): """ NOTE: This getter chunks the data incorrectly """ def __init__(self, max_order): self.max_order = max_order
[docs] def get_shape(self, basis): o = self.max_order return basis.nbf(), basis.nbf(), int((o+1)*(o+2)*(o+3)/6 - 1)
def __call__(self, out, factory, basis): intobj = factory.ao_multipoles(self.max_order) nsh = basis.nshell() for ish, jsh in xproduct(nsh, nsh): ioff, joff = map(basis.shell_to_basis_function, (ish, jsh)) inbf, jnbf = map(lambda sh: basis.shell(sh).nfunction, (ish, jsh)) intobj.compute_shell(ish, jsh) out.value[ioff:ioff+inbf, joff:joff+jnbf,:] = np.array(intobj.py_buffer).reshape( (inbf,jnbf,out.value.shape[2]) )
[docs]class AOOEIGetter(OEIGetter, MatrixGetter): VALID_KERNELS = [ "overlap", "kinetic", "potential" ]
[docs] def get_shape(self, basis): nbf = basis.nbf() return nbf, nbf
def __call__(self, out, mints): mints_getter = getattr(mints, "ao_" + self.kernel) psimat = mints_getter() self._copy_in_matrix(out, psimat) return out
[docs]class SOOEIGetter(OEIGetter, MatrixGetter): VALID_KERNELS = [ "overlap", "kinetic", "potential" ]
[docs] def get_shape(self, reference_wavefunction): nso = reference_wavefunction.nso() return nso, nso
def __call__(self, out, mints): mints_getter = getattr(mints, "so_" + self.kernel) psimat = mints_getter() self._copy_in_matrix(out, psimat) return out
[docs]class AOTEIGetterWithComputer(MemmapArrayGetter): VALID_INTEGRAL_TYPES = [ "eri", "f12", "f12g12", "f12_double_commutator", "f12_squared", "erf_eri", "erf_complement_eri" ] def __init__(self, integral_type): self.integral_type = integral_type if self.integral_type not in type(self).VALID_INTEGRAL_TYPES: raise ValueError("Don't know how to compute integrals of type '{0}'".format( integral_type )) # Make a special needs for the ERI doesn't include the correlation factor if self.integral_type == "eri": self.needs = [n for n in self.needs if n != "correlation_factor"] self.zero_basis_set = lambda: psi4.BasisSet.zero_ao_basis_set()
[docs] def get_shape(self, basis): nbf = basis.nbf() return nbf, nbf, nbf, nbf
def __call__(self, out, basis, correlation_factor=None): self._compute_ints(out, basis_sets=(basis, basis, basis, basis), correlation_factor=correlation_factor ) @staticmethod
[docs] def shell_slice(bs, ish): start = bs.shell_to_basis_function(ish) nbf = bs.shell(ish).nfunction return slice(start, start + nbf)
def _compute_ints(self, out, basis_sets, correlation_factor=None, clip_zero_basis=True): # Only create the zero_basis_set instance once we get here and need it if isinstance(self.zero_basis_set, lambda_type): self.zero_basis_set = self.zero_basis_set() #----------------------------------------# #region | Make the TwoBodyInt object | factory = psi4.IntegralFactory(*basis_sets) computer = getattr(factory, self.integral_type) if "f12" in self.integral_type: if correlation_factor is None: raise NotImplementedError() computer = computer(correlation_factor) elif "erf" in self.integral_type: raise NotImplementedError() else: computer = computer() #endregion #----------------------------------------# #region | Compute the integrals and store them in out.value | computer.set_enable_pybuffer(True) pybuffer = computer.py_buffer_object for shell_nums in xproduct(*[bs.nshell() for bs in basis_sets]): computer.compute_shell(*shell_nums) slices = tuple( AOTEIGetterWithComputer.shell_slice(bs, ish) for bs, ish in zip(basis_sets, shell_nums) if bs is not self.zero_basis_set ) buff = np.array(pybuffer) if clip_zero_basis: squeeze_axes =tuple(i for i, bs in enumerate(basis_sets) if bs is self.zero_basis_set) out.value[slices] = buff.squeeze(axis=squeeze_axes) else: out.value[slices] = buff #endregion #----------------------------------------# return
[docs]class ThreeCenterOverlapGetter(MemmapArrayGetter): def __init__(self): self.zero_basis_set = lambda: psi4.BasisSet.zero_ao_basis_set() def __call__(self, out, basis): # Only create the zero_basis_set instance once we get here and need it if isinstance(self.zero_basis_set, lambda_type): self.zero_basis_set = self.zero_basis_set() self._compute_ints(out, (basis, basis, basis, self.zero_basis_set)) def _compute_ints(self, out, basis_sets): factory = psi4.IntegralFactory(*basis_sets) computer = factory.overlap_3c() computer.set_enable_pybuffer(True) pybuffer = computer.py_buffer_object # Leave out the zero basis set for shell_nums in xproduct(*[bs.nshell() for bs in basis_sets[:-1]]): computer.compute_shell(*shell_nums) slices = tuple( AOTEIGetterWithComputer.shell_slice(bs, ish) for bs, ish in zip(basis_sets, shell_nums) if bs is not self.zero_basis_set ) buff = np.array(pybuffer) out.value[slices] = buff return
[docs]class DFThreeCenterAOTEIGetter(AOTEIGetterWithComputer): def __init__(self, two_center_bra, integral_type='eri'): self.two_center_bra = two_center_bra super(DFThreeCenterAOTEIGetter, self).__init__( integral_type=integral_type ) # noinspection PyMethodOverriding
[docs] def get_shape(self, basis, df_basis): nbf = basis.nbf() dfnbf = df_basis.nbf() if self.two_center_bra: return nbf, nbf, dfnbf else: return dfnbf, nbf, nbf # noinspection PyMethodOverriding
def __call__(self, out, basis, df_basis, correlation_factor=None): if isinstance(self.zero_basis_set, lambda_type): self.zero_basis_set = self.zero_basis_set() if self.two_center_bra: basis_sets = (basis, basis, df_basis, self.zero_basis_set) else: basis_sets = (df_basis, self.zero_basis_set, basis, basis) self._compute_ints(out, basis_sets, correlation_factor)
[docs]class DFTwoCenterAOTEIGetter(AOTEIGetterWithComputer):
[docs] def get_shape(self, df_basis): dfnbf = df_basis.nbf() return dfnbf, dfnbf
def __call__(self, out, df_basis, correlation_factor=None): if isinstance(self.zero_basis_set, lambda_type): self.zero_basis_set = self.zero_basis_set() self._compute_ints(out, basis_sets=(df_basis, self.zero_basis_set, df_basis, self.zero_basis_set), correlation_factor=correlation_factor )
[docs]class ArbitraryBasisDatumGetter(DatumGetter): def __init__(self, set_names): self.basis_set_names = tuple(bsname.lower() for bsname in set_names) @abstractproperty
[docs] def name(self): return NotImplemented
[docs]class ArbitraryBasisThreeCenterOverlapGetter(ArbitraryBasisDatumGetter, ThreeCenterOverlapGetter): def __init__(self, bs1, bs2, bs3): super(ArbitraryBasisThreeCenterOverlapGetter, self).__init__((bs1, bs2, bs3)) ThreeCenterOverlapGetter.__init__(self) @property
[docs] def name(self): return ( "ao__" + "__".join(self.basis_set_names[:2]) + "___overlap___" + "__".join(self.basis_set_names[2:]) )
[docs] def get_shape(self, basis_sets): return tuple(bs.nbf() for bs in basis_sets)
def __call__(self, out, basis_sets): # Only create the zero_basis_set instance once we get here and need it if isinstance(self.zero_basis_set, lambda_type): self.zero_basis_set = self.zero_basis_set() self._compute_ints(out, list(basis_sets) + [self.zero_basis_set])
[docs]class ArbitraryBasisAOTEIGetter(ArbitraryBasisDatumGetter, AOTEIGetterWithComputer): def __init__(self, integral_type, bs1, bs2=None, bs3=None, bs4=None): if bs2 is None: bs2 = bs1 if bs3 is None: bs3 = bs1 if bs4 is None: bs4 = bs2 super(ArbitraryBasisAOTEIGetter, self).__init__((bs1, bs2, bs3, bs4)) AOTEIGetterWithComputer.__init__(self, integral_type) @property
[docs] def name(self): return ( "ao__" + "__".join(self.basis_set_names[:2]) + "___" + self.integral_type + "___" + "__".join(self.basis_set_names[2:]) )
[docs] def get_shape(self, basis_sets): return tuple(bs.nbf() for bs in basis_sets)
def __call__(self, out, basis_sets, correlation_factor=None): self._compute_ints(out, basis_sets, correlation_factor)
[docs]class GeneralCorrelationFactorAOTEIGetter(ArbitraryBasisAOTEIGetter): def __init__(self, coefficients, exponents, *args, **kwargs): super(GeneralCorrelationFactorAOTEIGetter, self).__init__(*args, **kwargs) self.coefficients = coefficients self.exponents = exponents # noinspection PyMethodOverriding def __call__(self, out, basis_sets): cf = psi4.CorrelationFactor(len(self.coefficients)) coefs = psi4.Vector(len(self.coefficients)) expons = psi4.Vector(len(self.exponents)) for (ic, c), (ie, e) in zip(*map(enumerate, (self.coefficients, self.exponents))): coefs.set(0, ic, c) expons.set(0, ie, e) cf.set_params(coefs, expons) self._compute_ints(out, basis_sets, cf)
[docs]class AOOEIGetterWithComputer(MemmapArrayGetter): VALID_INTEGRAL_TYPES = [ "overlap", "kinetic", "potential" ] def __init__(self, integral_type): self.integral_type = integral_type if self.integral_type not in type(self).VALID_INTEGRAL_TYPES: raise ValueError("Don't know how to compute integrals of type '{0}'".format( integral_type )) # TODO this is very similar to the _compute_ints in AOTEIGetterWithComputer and could be combined in a common superclass def _compute_ints(self, out, basis_sets): #----------------------------------------# #region | Make the TwoBodyInt object | factory = psi4.IntegralFactory(*basis_sets) computer = getattr(factory, "ao_" + self.integral_type)() #endregion #----------------------------------------# #region | Compute the integrals and store them in out.value | computer.set_enable_pybuffer(True) pybuffer = computer.py_buffer_object for shell_nums in xproduct(*[bs.nshell() for bs in basis_sets]): computer.compute_shell(*shell_nums) slices = tuple( AOTEIGetterWithComputer.shell_slice(bs, ish) for bs, ish in zip(basis_sets, shell_nums) ) out.value[slices] = np.array(pybuffer) #endregion #----------------------------------------# return
[docs] def get_shape(self, basis): nbf = basis.nbf() return nbf, nbf
def __call__(self, out, basis, correlation_factor=None): self._compute_ints(out, basis_sets=(basis, basis) )
[docs]class ArbitraryBasisAOOEIGetter(ArbitraryBasisDatumGetter, AOOEIGetterWithComputer): def __init__(self, integral_type, bs1, bs2=None): if bs2 is None: bs2 = bs1 super(ArbitraryBasisAOOEIGetter, self).__init__((bs1, bs2)) AOOEIGetterWithComputer.__init__(self, integral_type) @property
[docs] def name(self): return "ao__" + ("___" + self.integral_type + "___").join(self.basis_set_names)
[docs] def get_shape(self, basis_sets): return tuple(bs.nbf() for bs in basis_sets)
def __call__(self, out, basis_sets): self._compute_ints(out, basis_sets)
[docs]class MOEigenvaluesGetter(MemmapArrayGetter): #TODO open shell #TODO symmetry
[docs] def get_shape(self, reference_wavefunction): return reference_wavefunction.nmo(),
def __call__(self, out, reference_wavefunction): vect = reference_wavefunction.epsilon_a() nbf = out.value.shape[0] for ibf in range(nbf): out.value[ibf] = vect[0, ibf] #endregion #================================================================================#
[docs]class ComputationCache(object): """ A set of computations cached in a given directory """ #========================================# #region | Class Attributes | # Always add new attributes to the end of these lists, # so that older files attributes can be seen as a # subset of newer files attributes (though right now # this ordering property isn't used) # Note that "molecule" is understood to be needed required_attributes = [ 'basis' ] available_psi_options = [ "scf_convergence" ] # Sentinel value for optional attributes to not include in the # key unless they are set by the user. NoDefaultValue = "___NoDefaultValue___" optional_attributes = AliasedDict({ 'correlation_factor_exponent' : 1.0, ('df_basis', 'df_basis_name', 'density_fitting_basis', 'fitting_basis', 'dfbasis') : NoDefaultValue }) case_sensative_attributes = [ ] do_writeback_on_sync = False #========================================# #region | Initialization | def __init__(self, directory, make_directory=False, make_path=False, make_shelve=True, writeback=None, read_only=False ): #========================================# #region Check for the directory and make it if needed. if writeback is None: writeback = not read_only if writeback and read_only: raise ValueError("Writeback cache can't be read only") self.directory = str(directory).rstrip(path_sep) self.read_only = read_only pth, dir_name = path_split(self.directory) if not isdir(pth): if exists(pth): raise OSError("File '{0}' exists, but it is not a directory.".format(pth)) if make_path: makedirs(pth) else: raise OSError("Path '{0}' does not exist.".format(pth)) if not isdir(self.directory): if make_directory: mkdir(self.directory) else: raise OSError("Directory {0} does not exist.".format(self.directory)) #endregion #----------------------------------------# #region Open the Shelf if read_only: flag = 'r' else: flag = 'c' if make_shelve else 'w' self.shelf = shelve.open( path_join(self.directory, "computation_shelf.db"), flag=flag, protocol=2, writeback=writeback ) self.analogous_keys = shelve.open( path_join(self.directory, "analogous_keys.db"), flag=flag, protocol=2, writeback=writeback ) self.writeback = writeback def close_shelf(shelf, other_shelf): try: shelf.close() except ValueError: # If it's already closed, it's okay pass try: other_shelf.close() except ValueError: # If it's already closed, it's okay pass if not self.read_only: atexit.register(close_shelf, self.shelf, self.analogous_keys) #endregion #----------------------------------------# #endregion #========================================# #region | Special Methods | def __iter__(self): for k in self.shelf: yield k def __getitem__(self, item): return self._get_existing_computation(item) def __setitem__(self, item, value): self.shelf[item] = value #endregion #========================================# #region | Methods |
[docs] def get_computation(self, molecule, needed_data=None, **kwargs ): """ @param molecule: @type molecule: Molecule, str, or anything for which Molecule(molecule) is sensible @param needed_data: @type needed_data: NoneType, Iterable @rtype: CachedComputation """ molkey = self.get_molecule_key(molecule) key = [molkey] key_mandatory = [molkey] optional_part = [] init_kwargs = dict(molecule=molecule) psi_options = dict() #----------------------------------------# #region | Gather required attributes | for arg in self.required_attributes: has_arg = arg in kwargs val = kwargs.pop(arg, None) if not has_arg: # Cheesy case insensitivity found = False for kw in kwargs: if kw.lower() == arg.lower(): val = kwargs.pop(kw) found = True break if not found: raise ValueError("Missing required argument {0}".format(arg)) if isinstance(val, str) and arg not in ComputationCache.case_sensative_attributes: val=val.lower() key.append((arg,val)) key_mandatory.append((arg,val)) init_kwargs[arg] = val #endregion #----------------------------------------# #region | Gather optional attributes | # This isn't quite perfect, since data that do not depend on # a given optional attribute will be recomputed when really # that data could be reused in a new computation which has # some extra optional argument. I don't see an easy way # around this right now, though; remember that iterating # over existing Computations should be out of the question. optional_args = dict() for keyset, defaultval in self.optional_attributes.items(): kw = [kw for kw in kwargs if kw in keyset] kw = None if len(kw) == 0 else kw[-1] val = kwargs.pop(kw, None) has_arg = kw is not None if not has_arg: # Cheesy case insensitivity lower_keyset = set(k.lower() for k in keyset) for kw in kwargs: if kw.lower() in lower_keyset: val = kwargs.pop(kw) has_arg = True break firstkey = self.optional_attributes.firstkey(keyset) if isinstance(val, str) and firstkey not in ComputationCache.case_sensative_attributes: val=val.lower() if has_arg: key.append((firstkey,val)) optional_part.append((firstkey,val)) optional_args[firstkey] = val elif not defaultval == ComputationCache.NoDefaultValue: key.append((firstkey,val)) optional_part.append((firstkey,val)) optional_args[firstkey] = defaultval if len(kwargs) > 0: raise ValueError("Unknown or duplicate attribute {0}".format(kwargs.keys()[0])) #endregion #----------------------------------------# #region | Gather psi options | # These get set before each UnitializedDependency is evaluated and restored to their # previous values after the given uninitialized dependency is evaluated. Options # that carry around baggage should not be used with this (for instance, 'basis' # needs for the global environment to have a molecule defined or it raises an # exception) for arg in self.available_psi_options: has_arg = arg in kwargs val = kwargs.pop(arg, None) if not has_arg: # Cheesy case insensitivity for kw in kwargs: if kw.lower() == arg.lower(): val = kwargs.pop(kw) has_arg = True break if isinstance(val, str) and arg not in ComputationCache.case_sensative_attributes: val=val.lower() if has_arg: key.append((arg,val)) key_mandatory.append((arg,val)) psi_options[arg] = val #endregion #----------------------------------------# # This is not ideal. For instance, two computations # with scf_convergence set to 8 and 8.0 would show # up as unique. But shelve apperently requires a # string to hash with. key_tuple = tuple(key) key = str(key_tuple) key_mandatory = str(tuple(key_mandatory)) if key not in self.shelf: # The directory name isn't all that important, # since it won't (and shouldn't) be human # readable anyway. The complicated hashing # is just to get a unique key. # The abs() is because I don't like directory # names that start with "-" hash_key = abs(hash(key)) hash_dir = "{0:016x}".format(hash_key)[:16] while exists(path_join(self.directory, hash_dir)): # Just choose the next available value. This # should pretty much never happen, since # an existing directory should also have # an existing key in the shelf. But just in # case... hash_key += 1 hash_dir = "{0:016x}".format(hash_key)[:16] hash_dir = path_join(self.directory, hash_dir) # Now make the CachedComputation object. No need # to sync afterwards, since the assignment triggers # an automatic sync. rv = CachedComputation( needed_data=needed_data, directory=hash_dir, owner=self, optional_arguments=optional_args, **init_kwargs ) rv.shelf_key = key rv.minimal_shelf_key = key_mandatory # Add the key to the list of keys related to the # mandatory minimal subset of keys self._map_analogous_keys(key_mandatory, key_tuple) # Make sure the object is constructed successfully # before shelving it self.shelf[key] = rv else: rv = self._get_existing_computation(key, key_mandatory) # Add the key to the list of keys related to the # mandatory minimal subset of keys if it isn't # already there self._map_analogous_keys(key_mandatory, key_tuple) if needed_data is not None: rv.get_data(needed_data) return rv
[docs] def sync_computation(self, comp, sync_writeback=None): self.shelf[comp.shelf_key] = comp if ((sync_writeback is None and self.do_writeback_on_sync) or sync_writeback) and self.writeback: self.shelf.sync()
[docs] def clear_datum(self, datum_name, fail_if_missing=False): """ Clear datum named `datum_name` from all computations in the cache @param datum_name: The name of the datum to clear from all known computations @type datum_name: str @param fail_if_missing: Whether or not to raise an exception if no computations have a datum named `datum_name` @return: The number of data deleted @rtype: int @see CachedComputation.clear_datum, clear_data_regex """ num_found = 0 for key in self.shelf: comp = self._get_existing_computation(key) cleared = comp.clear_datum(datum_name) if cleared: num_found += 1 if fail_if_missing and num_found == 0: raise ValueError("No known computations have a datum named '{0}'".format(datum_name)) return num_found
[docs] def clear_data_regex(self, regex, fail_if_missing=False, flags=0): """ Clear all data matching regex from known computations @param regex: The regular expression to search for in the name of the datum @type regex: str or compiled regex @param fail_if_missing: Whether or not to raise an exception if no data match the regex @param flags: Flags to compile the regular expression with. Ignored if regex is an instance of re.RegexObject. @type flags: int @raise ValueError: if no data names match the regex and fail_if_missing is True @return: Length 2 tuple of the number of data removed and the number of computations it was removed from. @rtype: tuple @see CachedComputation.clear_data_regex """ num_found = 0 num_comps = 0 if hasattr(regex, "pattern") and hasattr(regex, "search"): re_string = regex.pattern else: re_string = str(regex) regex = re.compile(re_string, flags) for key in self.shelf: comp = self._get_existing_computation(key) found = comp.clear_data_regex(regex, flags=flags) if found > 0: num_found += found num_comps += 1 if fail_if_missing and num_found == 0: raise ValueError("No known computations have a datum matching '{0}'".format(re_string)) return num_found, num_comps #endregion #========================================# #region | Private Methods |
def _get_existing_computation(self, key, key_mandatory=None): """ Internal use only @type key: str @type key_mandatory: str @rtype : CachedComputation @raise KeyError : if key is not found in the computation """ rv = self.shelf[key] rv.shelf_key = key if key_mandatory is not None: rv.minimal_shelf_key = key_mandatory rv.owner = self return rv def _map_analogous_keys(self, key_mandatory, key_tuple): if key_mandatory in self.analogous_keys: # Don't append, since mutation of the object will # not get immediately registered. Instead, assign # the appended list known_keys = self.analogous_keys[key_mandatory] if key_tuple not in known_keys: self.analogous_keys[key_mandatory] = self.analogous_keys[key_mandatory] + [key_tuple] else: self.analogous_keys[key_mandatory] = [key_tuple] #endregion #========================================# #region | Class Methods | @classmethod
[docs] def get_molecule_key(cls, molecule): """ For now, this just returns the xyz string representation of the Molecule object. A more sophisticated means could be devised if constant recomputation becomes a problem, but it would be a total mess (see my attempts in the MoleculeStub class in grendel) """ if not isinstance(molecule, Molecule): return Molecule(molecule).xyz_string(header=False) else: return molecule.xyz_string(header=False) #endregion #========================================# # End of ComputationCache class
pass
[docs]class CachedComputation(object): """ A computation on a given molecule with various relevant integrals and such cached for later as tensors or numpy memory maps. """ #========================================# #region | Class Attributes | PICKLE_VERSION = (2,1,2) allow_analogous_load_ever = True # For now, set a hard max of 4 GB default_memory = min(resource.getrlimit(resource.RLIMIT_RSS)[0], 4*1024*1024*1024) if default_memory <= 0: # Just set to 4 GB and hope... default_memory = 4*1024*1024*1024 _statics_initialized = False @classmethod def _init_statics(cls): """ Initialize static variables. This gets called the first time a CachedComputation object is constructed. Wrapping initialization of various things in this class method keeps them from being initialized when the module is loaded, which can be a problem in some cases (particularly when generating documentation) """ #----------------------------------------# # Initialize Psi4 Memory to a reasonable value try: psi4.be_quiet() psi4.set_memory(cls.default_memory) finally: psi4.reopen_outfile() #----------------------------------------# # Initialize the basis set parser cls.parser = psi4.Gaussian94BasisSetParser() #----------------------------------------# array_getters = dict() # Add the AO and MO TEI getters to the known getters for kernel in TEIGetter.VALID_KERNELS: array_getters["ao_" + kernel] = AOTEIGetter(kernel) array_getters["mo_" + kernel] = MOTEIGetter(kernel) for kernel in AOOEIGetter.VALID_KERNELS: array_getters["ao_" + kernel] = AOOEIGetter(kernel) # Add the MO Coefficient getters for orbital_type in MOCoefficientsGetter.ALLOWED_TYPES: if orbital_type == "canonical": array_getters["mo_coefficients"] = MOCoefficientsGetter("canonical") array_getters["occupied_mo_coefficients"] = OccupiedMOCoefficientsGetter("canonical") else: array_getters["mo_coefficients" + "_" + orbital_type] = MOCoefficientsGetter(orbital_type) array_getters["occupied_mo_coefficients" + "_" + orbital_type] = OccupiedMOCoefficientsGetter(orbital_type) array_getters["mo_eigenvalues"] = MOEigenvaluesGetter() other_getters = dict() # Simple getters that amount to nothing more than # a method call on a single dependency # SimpleMethodCallGetters for molecule _molecule_methods = ["nuclear_repulsion_energy", "natom", "name"] for method in _molecule_methods: other_getters["molecule_" + method] = SimpleMethodCallGetter('psi_molecule', method) # SimpleMethodCallGetters for basis _basis_methods = ["nbf", "nao", "nprimitive", "nshell", "has_puream"] for method in _basis_methods: other_getters["basis_" + method] = SimpleMethodCallGetter('basis', method) # SimpleMethodCallGetters for wavefunction _reference_wavefunction_methods = [ "nso", "nmo", "nirrep", "energy", "nalpha", "nbeta" ] for method in _reference_wavefunction_methods: other_getters["reference_wavefunction_" + method] = SimpleMethodCallGetter('reference_wavefunction', method) _reference_wavefunction_dimension_methods = [ "doccpi", "soccpi", "nsopi", "nalphapi", "nbetapi", "frzcpi", "frzvpi" ] for method in _reference_wavefunction_dimension_methods: other_getters["reference_wavefunction_" + method] = DimensionListGetter('reference_wavefunction', method) default_psi_options = dict( scf_type = "direct" ) #endregion #========================================# #region | Initialization | @typecheck( molecule=(Molecule, str), cached_data=(dict, None) ) def __init__(self, molecule, basis, directory, optional_arguments, needed_data=None, cached_data=None, owner=None, psi_options=None, ): #========================================# # only initialize static class variables once if not CachedComputation._statics_initialized: CachedComputation._init_statics() CachedComputation._statics_initialized = True #========================================# #region Set up arguments and attributes if isinstance(molecule, str): self.molecule = Molecule(molecule) else: self.molecule = molecule self.basis_name = str(basis) if needed_data is None: needed_data = [] elif isinstance(needed_data, str): needed_data = [needed_data] elif isinstance(needed_data, Iterable): needed_data = list(needed_data) self.directory = directory if not isdir(self.directory): if exists(self.directory): raise IOError("File {0} exists, but it is not a directory.".format(self.directory)) else: mkdir(self.directory) self.owner = owner self.optional_arguments = ComputationCache.optional_attributes.copy() self.optional_arguments.update(optional_arguments) if any(k not in ComputationCache.optional_attributes for k in self.optional_arguments): raise ValueError("Invalid optional_arguments key: {0}".format( tuple(next(k for k in self.optional_arguments if k not in ComputationCache.optional_attributes)) )) self._custom_getters = dict() self._basis_registry = AliasedDict() self._optional_argument_dependencies = dict() self._analogously_loaded_data = set() self.parallel_ready = False self._dependency_aliases = defaultdict(lambda: []) # multiprocessing RLock object for locking # psi to avoid PSIO errors self.psi_lock = None self._required_locks = dict() self._original_names = dict() self.allow_analogous_load = True #endregion #========================================# #region psi options # Psi4 options get set before the evaluation # of any lazy attributes and set back to there # former values after that evaluation. This # prevents a change of option in one computation # from affecting the option value in another # computation which assumes the default. self._psi_options = CachedComputation.default_psi_options.copy() if psi_options is not None: self._psi_options.update(psi_options) #endregion #========================================# #region Set up the dictionary of data bits to cache if cached_data is None: self.cached_data = dict() else: self.cached_data = cached_data #endregion #========================================# #region Set up the psi molecule and psi objects # These objects don't get pickled #----------------------------------------# # Lazily evaluated attributes self.register_dependency( "psi_molecule", UninitializedDependency( # Don't use symmetry for now # Note that psi4.geometry calls activate on the molecule lambda self: psi4.geometry("symmetry c1\n" + self.molecule.xyz_string(header=False).strip()), self, depends_on=[] ) ) # basis, df_basis, and add the unit basis (a.k.a. zero basis) to the registry self.register_dependency( "basis", self.uninitialized_basis(self.basis_name) ) self._basis_registry[self.basis_name.lower()] = self.basis self.register_dependency( "df_basis", UninitializedDependency( partial( CachedComputation.construct_basis, basis_name=self.optional_arguments['df_basis'] ), self, depends_on=['psi_molecule'], requires_optional_arguments=['df_basis'] ), dependent_optional_arguments=['df_basis'] ) if self.optional_argument_given("df_basis"): self._basis_registry[self.optional_arguments['df_basis'].lower()] = self.df_basis self._basis_registry[('unit','1','zero','one')] = UninitializedDependency( CachedComputation.construct_unit_basis, self, depends_on=['psi_molecule'] ) # other stuff self.register_dependency( "mints", UninitializedDependency( lambda self: psi4.MintsHelper(self.basis), self, depends_on=['basis'] ) ) self.register_dependency( "factory", UninitializedDependency( lambda self: self.mints.integral(), self, depends_on=['mints'] ) ) def get_scf_wavefunction(self): # Check the basis to be sure it's the right one if psi4.options.basis.lower() != self.basis_name.lower(): psi4.options.basis = self.basis_name # run SCF scratch_path = tempfile.mkdtemp() psi4.IOManager.shared_object().set_default_path(scratch_path) psi4.energy("scf") psi4.clean() shutil.rmtree(scratch_path) # grab the wavefunction of the most recently run computation (scf, in this case) return psi4.wavefunction() self.register_dependency( "reference_wavefunction", UninitializedDependency( get_scf_wavefunction, self, depends_on=['basis'] ), required_locks=[], #required_locks=["psi_lock"] ) self.register_dependency( "correlation_factor", UninitializedDependency( lambda self: psi4.FittedSlaterCorrelationFactor( self.optional_arguments['correlation_factor_exponent'] ), self, depends_on=['psi_molecule'] ), dependent_optional_arguments=['correlation_factor_exponent'] ) # TODO handle open shells self.register_dependency( "docc_space", UninitializedDependency( lambda self: self.reference_wavefunction.alpha_orbital_space('i', 'SO', 'OCC'), self, depends_on=['reference_wavefunction'] ), aliases=["doubly_occupied_space", "occupied_space", "alpha_occupied_space"] ) self.register_dependency( "virt_space", UninitializedDependency( lambda self: self.reference_wavefunction.alpha_orbital_space('a', 'SO', 'VIR'), self, depends_on=['reference_wavefunction'] ), aliases=["virtual_space", "alpha_virtual_space"] ) self.register_dependency( "mo_coefficients", UninitializedDependency( lambda self: self.reference_wavefunction.Ca(), self, depends_on=['reference_wavefunction'] ), aliases=["mo_coefficients_canonical", "canonical_mo_coefficients"] ) #endregion #========================================# #region Get the data we need while len(needed_data) != 0: needed_name = needed_data.pop(0) self._fill_datum(needed_name) #endregion #========================================# #endregion #========================================# @classmethod
[docs] def get_nbasis(cls, mol, basis): directory = tempfile.gettempdir() cc = CachedComputation(mol, basis, directory, {}) return cc.get_lazy_attribute("basis").nbf() #========================================# #region | Methods |
[docs] def optional_argument_given(self, arg): return self.optional_arguments[arg] != ComputationCache.NoDefaultValue
[docs] def construct_basis(self, basis_name=None): if basis_name is None: basis_name = self.basis_name old_value = psi4.options.basis psi4.options.basis = basis_name rv = psi4.BasisSet.construct(CachedComputation.parser, self.psi_molecule, 'BASIS') if old_value != '': psi4.options.basis = old_value self._basis_registry[basis_name.lower()] = rv return rv
[docs] def construct_unit_basis(self): rv = psi4.BasisSet.zero_ao_basis_set() self._basis_registry['unit'] = rv return rv
[docs] def uninitialized_basis(self, basis_name): return UninitializedDependency( partial(CachedComputation.construct_basis, basis_name=basis_name), self, depends_on=['psi_molecule'] )
[docs] def register_dependency( self, attribute_name, dependency_object_or_function, depends_on=None, aliases=None, dependent_optional_arguments=None, required_locks=tuple() ): if isinstance(dependency_object_or_function, UninitializedDependency): dependency_object = dependency_object_or_function if depends_on is not None: dependency_object.depends_on.extend(depends_on) else: if depends_on is None: depends_on = [] dependency_object = UninitializedDependency( dependency_object_or_function, self, depends_on ) setattr(self, attribute_name, dependency_object) if aliases is not None: # Register pointers to the original object under alternate names for alias in aliases: self.register_dependency_alias(alias, attribute_name) if dependent_optional_arguments is not None: self._optional_argument_dependencies[attribute_name] = dependent_optional_arguments else: self._optional_argument_dependencies[attribute_name] = [] self._required_locks[attribute_name] = list(required_locks)
[docs] def register_dependency_alias(self, alias, original_name): attr = getattr(self, original_name) self._dependency_aliases[attr].append(alias) self._original_names[alias] = original_name setattr(self, alias, attr)
[docs] def register_basis(self, basis_name): if basis_name not in self._basis_registry: self._basis_registry[basis_name.lower()] = self.uninitialized_basis(basis_name)
[docs] def get_datum(self, name=None, custom_getter=None, pre_computation_callback=None, post_computation_callback=None, analogous_load_callback=lambda akey, fname=None: print( "Loaded datum from analogous source file {0} with computation attributes {1}".format( "(no file)" if fname is None else fname, akey[1:] )), allow_analogous_load=None ): if allow_analogous_load is None: allow_analogous_load = self.allow_analogous_load if not CachedComputation.allow_analogous_load_ever: allow_analogous_load = False #----------------------------------------# if name is None: if not hasattr(custom_getter, "name"): raise TypeError() name = custom_getter.name #----------------------------------------# if not self.has_datum(name): if (name in self._custom_getters and name in self.cached_data and not self.cached_data[name].filled ): del self._custom_getters[name] self._fill_datum( name, custom_getter, pre_computation_callback=pre_computation_callback, post_computation_callback=post_computation_callback, allow_analogous_load=allow_analogous_load, analogous_load_callback=analogous_load_callback ) # Sync the parent shelf since self has been modified if self.owner is not None and not self.parallel_ready: self.owner.sync_computation(self) return self.cached_data[name]
[docs] def get_data(self, names): rv = [] for name in names: rv.append(self.get_datum(name)) return rv
[docs] def clear_datum(self, name, fail_if_missing=False, clear_analogous=False): """ Clear a datum from the computation. If the datum has a file associated with it, delete the file. @param name: The name of the datum to clear @type name: str @param fail_if_missing: Whether or not to raise an exception if the datum is not found @type fail_if_missing: bool @raise ValueError: if the datum named `name` is not found @return: True if the datum was found and deleted, False otherwise @rtype: bool """ if name in self.cached_data: if name not in self._analogously_loaded_data or clear_analogous: # In the clear_analogous case, the analogous computation will see the data # as corrupted next time it is loaded. This is probably what we want in # this case anyway d = self.cached_data[name] if isinstance(d, CachedMemmapArray): fname = d.filename del self.cached_data[name] if exists(fname): os.remove(fname) else: raise_warning( "File '{0}' associated with datum named '{1}' is missing and" " cannot be deleted".format( fname, name ), FileMissingWarning ) else: del self.cached_data[name] else: # Just remove the name from the cache del self.cached_data[name] self._analogously_loaded_data.remove(name) if name in self._custom_getters: del self._custom_getters[name] # Sync the parent shelf since self has been modified if self.owner is not None and not self.parallel_ready: self.owner.sync_computation(self) return True elif fail_if_missing: raise ValueError("Datum '{0}' does not exist. Known data names are:\n{1}".format( name, " " + "\n ".join(sorted(self.cached_data.keys())) )) return False
[docs] def clear_data_regex(self, regex, fail_if_missing=False, flags=0, clear_analogous=False): """ Clear all data matching regex @param regex: The regular expression to search for in the name of the datum @type regex: str or re.RegexObject @param fail_if_missing: Whether or not to raise an exception if no data match the regex @param flags: Flags to compile the regular expression with. Ignored if regex is an instance of re.RegexObject. @type flags: int @raise ValueError: if no data names match the regex and fail_if_missing is True @return: The number of data removed @rtype: int """ found_count = 0 if hasattr(regex, "pattern") and hasattr(regex, "search"): re_string = regex.pattern else: re_string = str(regex) regex = re.compile(re_string, flags) for name in list(self.cached_data.keys()): if regex.search(name) is not None: found_count += 1 self.clear_datum(name, clear_analogous=clear_analogous) if fail_if_missing and found_count == 0: raise ValueError("No data matching '{0}' found. Known data names are:\n{1}".format( re_string, " " + "\n ".join(self.cached_data.keys()) )) if found_count > 0: if self.owner is not None and not self.parallel_ready: self.owner.sync_computation(self) return found_count
[docs] def clear_all_data(self, clear_analogous=False): for datum_name in list(self.cached_data.keys()): self.clear_datum(datum_name, clear_analogous)
[docs] def get_lazy_attribute(self, needed_attr): # Make sure we know how to get the attribute if not hasattr(self, needed_attr): raise NameError("Don't know how to get attribute '{0}'".format(needed_attr)) #----------------------------------------# if needed_attr in self._original_names: needed_attr = self._original_names[needed_attr] acquired = [] try: for lock_name in self._required_locks[needed_attr]: lock = getattr(self, lock_name) if lock is not None: lock.acquire(blocking=True) acquired.append(lock) elif self.parallel_ready: raise ValueError("Required lock named '{}' not set for" " parallel ready CachedComputation".format( lock_name )) if self.parallel_ready: psi4.clean() self._psi_options_use() attr = getattr(self, needed_attr) if isinstance(attr, UninitializedDependency): aliases = self._dependency_aliases[attr] attr = attr.initialize() setattr(self, needed_attr, attr) for alias in aliases: setattr(self, alias, attr) if self.parallel_ready: psi4.clean() self._psi_options_restore() finally: for lock in acquired: lock.release() return attr
[docs] def get_file_path(self, datum_name): return CachedMemmapArray.normalized_name(path_join(self.directory, datum_name + ".npy"))
[docs] def store_datum(self, name, value): """ @param name: str @param value: object @raise: NameError """ if name in self.cached_data: if not isinstance(value, np.ndarray) and not self.cached_data[name].filled: if self.cached_data[name] == value: # Forgot to set the filled flag before self.cached_data[name].filled = True else: raise NameError("Already have cached datum named '{0}'".format(name)) elif name in self.array_getters or name in self.other_getters: raise NameError("Naming conflict: Data storage operation requested for" " datum named '{0}', but a getter for this datum is " " already known in CachedComputation.{1}".format( name, "array_getters" if name in self.array_getters else 'other_getters' )) elif name in self._custom_getters: raise NameError("Naming conflict: Custom getter for datum named '{0}'" " already exists".format( name )) #----------------------------------------# if isinstance(value, np.ndarray): out = CachedMemmapArray( filename=self.get_file_path(name), shape=value.shape, # Force overwrite because if the file exists at this point, # it's not attached to self.cached_data, so it's useless # anyway. force_overwrite=True ) out.value[...] = value out.filled = True self.cached_data[name] = out else: # Store it as a regular Datum out = CachedDatum(value) out.filled = True self.cached_data[name] = out #----------------------------------------# # Sync the parent shelf since self has been modified if self.owner is not None and not self.parallel_ready: self.owner.sync_computation(self)
[docs] def has_datum(self, name): return (name in self.cached_data and isinstance(self.cached_data[name], CachedDatum) and self.cached_data[name].filled )
[docs] def has_data(self, *args): for name in args: if not self.has_datum(name): return False return True
[docs] def save_computation(self): if self.parallel_ready: raise IOError("Cannot save computation during parallel computation") if self.owner is None: raise ValueError("Can't save computation because parent cache doesn't exist") self.owner.sync_computation(self) self.owner.shelf.sync()
[docs] def begin_parallel_use(self): self.parallel_ready = True
[docs] def end_parallel_use(self): self.parallel_ready = False
[docs] def set_psi_memory(self, value): psi4.set_memory(value) #endregion #========================================# #region | Special Methods |
def __reduce_ex__(self, protocol): unloader = CachedComputationUnloader(self) return unloader, tuple() def __del__(self): if self.owner is not None and not self.owner.read_only: self.owner.sync_computation(self) #endregion #========================================# #region | Private Methods | def _psi_options_use(self): if not isinstance(self.psi_molecule, UninitializedDependency): psi4.activate(self.psi_molecule) self._psi_saved_opts = dict() for opt, val in self._psi_options.items(): self._psi_saved_opts[opt] = getattr(psi4.options, opt) setattr(psi4.options, opt, val) def _psi_options_restore(self): for opt, val in self._psi_saved_opts.items(): setattr(psi4.options, opt, val) def _fill_datum(self, needed_name=None, getter=None, pre_computation_callback=None, post_computation_callback=None, analogous_load_callback=None, allow_analogous_load=True ): if needed_name is None: if not hasattr(getter, 'name'): raise TypeError() needed_name = getter.name #----------------------------------------# if "ao_multipole_" in needed_name and needed_name not in self.array_getters: max_order = int(needed_name[13:]) self.array_getters[needed_name] = AOMultipoleGetter(max_order) #----------------------------------------# if getter is not None: if needed_name in self.array_getters or needed_name in self.other_getters: raise NameError("Naming conflict: Custom getter requested for datum named" " '{0}', but a getter for this datum is already known in" " CachedComputation.{1}".format( needed_name, "array_getters" if needed_name in self.array_getters else 'other_getters' )) elif needed_name in self._custom_getters: raise NameError("Naming conflict: Custom getter for datum named '{0}'" " already exists".format( needed_name )) self._custom_getters[needed_name] = getter #========================================# # First check if we've already filled the given datum if needed_name in self.cached_data and self.cached_data[needed_name].filled: return #----------------------------------------# #region | Handle MemmapArrayGetters | if ((getter is not None and isinstance(getter, MemmapArrayGetter)) or needed_name in self.array_getters ): fname = self.get_file_path(needed_name) if getter is None: getter = self.array_getters[needed_name] shape, _, dtype = CachedMemmapArray.read_file_metadata(fname) load_successful = False # Try to load the memmap if we successfully retrieved a shape # from the cached file. has_tb = False if shape is not None: try: # The shape and dtype arguments present redundant information, # (which is retrieved when the file is loaded anyway) # but including them here does no harm, and they may be used # for error checking or something else in the future. out = CachedMemmapArray(fname, shape=shape, dtype=dtype ) fname = out.filename load_successful = True except ValueError: # Raised if the data is invalid. This means the file is corrupted # and needs to be rewritten load_successful = False out = None has_tb = True except IOError: # Raised if the file cannot be opened correctly. This should never # really happen unless the file changes between the call to # CachedMemmapArray.read_file_metadata() and here. load_successful = False out = None has_tb = True # Only compute the shape using the getter if we need to. if not load_successful: loaded_key = False # For now, if we are parallel ready, don't borrow data if allow_analogous_load and not self.parallel_ready: loaded_key = self._check_analogous_computation(needed_name, getter) if loaded_key is not False: # Another computatation has a compatible datum. We can use it. # The helper function has already put it in our cached_data # dictionary, so all we need to do is use it. out = self.cached_data[needed_name] fname = out.filename # We should check that there isn't a file in our own directory, # and if so, delete it. if exists(fname): raise_warning( "File '{0}' is corrupted (couldn't get shape), but an analogous" " datum was found in '{1}'; the corrupted file will be deleted" " and the analogous one will be used.".format( fname, out.filename ), FileOverwriteWarning ) os.remove(fname) self._analogously_loaded_data.add(needed_name) if callable(analogous_load_callback): analogous_load_callback(loaded_key, out.filename) else: # We don't have the file, or the file is invalid, # so we need to first get the shape and then # create the file. shape_needs = getargspec(getter.get_shape).args[1:] # exclude "self" shape_kwargs = self._fill_needed_kwargs(shape_needs, getter) shape = getter.get_shape(**shape_kwargs) # If the file exists but we've gotten here, this means that the file # somehow corrupted and needs to be overwritten. Warn the user # of the situation. if exists(fname): if has_tb: raise_warning( "File '{0}' is corrupted and will be overwritten." " Its data will be recomputed. Traceback:\n {1}".format( fname, "\n ".join(traceback.format_exc())), FileOverwriteWarning ) else: raise_warning( "File '{0}' is corrupted (couldn't get shape) and will be overwritten." " Its data will be recomputed.".format( fname), FileOverwriteWarning ) # Construct the CachedMemmapArray with force_overwrite=True since # we already checked for a valid file earlier and found that # we couldn't determine the shape. out = CachedMemmapArray( fname, shape=shape, force_overwrite=True ) #----------------------------------------# # Only fill it if we need to if not out.filled: if out.read_only: # This shouldn't happen raise_warning( "File '{0}' exists, but was not marked as filled previously. It's data" " will be recomputed.".format( fname ), FileOverwriteWarning ) out = CachedMemmapArray( fname, shape=shape, force_overwrite=True ) if pre_computation_callback is not None: pre_computation_callback() getter_kwargs = self._fill_needed_kwargs(getter.needs, getter) # This is where the computation of the data actually happens try: getter(out, **getter_kwargs) except Exception: # Remove the file caching the failure if exists(fname): remove(fname) raise else: out.filled = True if post_computation_callback is not None: post_computation_callback() self.cached_data[needed_name] = out #endregion #----------------------------------------# #region | Handle simpler getters | elif isinstance(getter, DatumGetter) or needed_name in CachedComputation.other_getters: if not isinstance(getter, DatumGetter): getter = CachedComputation.other_getters[needed_name] loaded_key = False # For now, if we are parallel ready, don't borrow data if allow_analogous_load and not self.parallel_ready: loaded_key = self._check_analogous_computation(needed_name, getter) if loaded_key is not False: self._analogously_loaded_data.add(needed_name) if callable(analogous_load_callback): analogous_load_callback(loaded_key) else: if pre_computation_callback is not None: pre_computation_callback() getter_kwargs = self._fill_needed_kwargs(getter.needs, getter) self.cached_data[needed_name] = CachedDatum(getter(**getter_kwargs)) self.cached_data[needed_name].filled = True if post_computation_callback is not None: post_computation_callback() #endregion #----------------------------------------# else: raise NotImplementedError("Don't know how to get '{}'".format(needed_name)) def _check_analogous_computation(self, needed_datum_name, getter): if self.owner is not None and hasattr(self, "minimal_shelf_key"): analogs = self.owner.analogous_keys[self.minimal_shelf_key] needed_opt = [] # remove the basis_sets attribute which is used by arbitrary basis getters # and does not serve to make one computation distinct from another anyway needs_no_bs = [g for g in getter.needs if g != "basis_sets"] for needed_attr in needs_no_bs: needed_opt.extend(self._optional_argument_dependencies[needed_attr]) for akey in analogs: if str(akey) == self.shelf_key: # Don't check ourselves... continue useable = True for opt in needed_opt: # Skip the molecule key using akey[1:]; it's never optional avalues = [v for k, v in akey[1:] if k == opt] if len(avalues) == 0: # A required optional argument was not given # for the analogou,s computation, so we # can't use it useable = False break elif len(avalues) == 2: # This should never happen raise KeyError("Something went horribly wrong; multiply defined key") else: # len(avalues) == 1 if avalues[0] != self.optional_arguments[opt]: # It doesn't have the same value, so we # can't use it useable = False break if useable: acomp = self.owner._get_existing_computation(str(akey), self.minimal_shelf_key) if acomp.has_datum(needed_datum_name): self.cached_data[needed_datum_name] = acomp.get_datum(needed_datum_name) return akey return False def _fill_needed_kwargs(self, needed_kws, getter): rv = dict() for needed_attr in needed_kws: if isinstance(getter, ArbitraryBasisDatumGetter) and needed_attr == "basis_sets": sets = [] for bsname in getter.basis_set_names: if bsname.lower() not in self._basis_registry: raise NameError("Unknown basis set '{0}'".format(bsname)) bs = self._basis_registry[bsname.lower()] if isinstance(bs, UninitializedDependency): bs = bs.initialize() if bsname == self.basis_name: self.basis = bs elif self.optional_argument_given('df_basis') and bsname == self.optional_arguments['df_basis']: self.df_basis = bs sets.append(bs) rv[needed_attr] = tuple(sets) else: rv[needed_attr] = self.get_lazy_attribute(needed_attr) return rv #endregion #========================================# # end CachedComputation class pass #================================================================================# #region | Cached data classes and unloaders |
[docs]class CachedDatum(object): """ Anything from a computation that might need to be cached. Assuming value is picklable, the pickling functionality of this class is trivial. Subclasses may override for more complex unpickling behavior """ def __init__(self, value): # Remember, this only gets called when constructed normally, not when unpickled self.value = value
[docs]class CachedMemmapArray(CachedDatum): """ An instance of a numpy.ndarray subclass to be cached. Only the filename is pickled """ #################### # Class Attributes # #################### DEFAULT_MMAP_LOAD_MODE = 'r' files_needing_flush = set() ################## # Initialization # ################## # noinspection PyMissingConstructor def __init__(self, filename, dtype="float64", shape=None, force_overwrite=False, mmap_load_mode=DEFAULT_MMAP_LOAD_MODE ): self.filename = CachedMemmapArray.normalized_name(filename) self.dtype = dtype # Note: mmap_load_mode is the mode used when loading. When creating, # "w+" is always used. self.mmap_load_mode=mmap_load_mode self.shape = shape self.write_mode = False self.force_overwrite = force_overwrite self.read_only = not force_overwrite and exists(self.filename) self.filled = False ################### # Special Methods # ################### def __getstate__(self): state = self.__dict__.copy() # Just get rid of the value, leave everything else if 'value' in state: # remember, value only gets filled if it is accessed del state['value'] return state def __setstate__(self, state): self.__dict__.update(state) # Note that the 'filled' attribute will stick with the class # from the pickled instance self.force_overwrite = False ################# # Class Methods # ################# @classmethod
[docs] def normalized_name(cls, filename): spl = path_split(filename) dirname, endname = path_join(*spl[:-1]), spl[-1] if len(endname) > 240: # The filename will be too long; we need to make a hash of the name and basically # hope for no collisions. If collisions happen (very low probability), # the data will just be overwritten from some other data value # Combine two hashing techniques to further avoid collisions hash_name = hashlib.sha224(endname).hexdigest() + hashlib.md5(endname).hexdigest() + ".npy" return path_join(dirname, hash_name) else: return filename
@classmethod
[docs] def read_file_metadata(cls, filename): """ Try to get the shape, ordering, and dtype of the memmap cached in the file `filename`. If the file doesn't exist, isn't readable, or the data is invalid, returns (None, None, None). See numpy.lib.format.read_array_header_1_0() """ try: fp = open(filename) major, minor = read_magic(fp) shape, order, dtype = read_array_header_1_0(fp) fp.close() return shape, order, dtype except ValueError: return None, None, None except IOError: return None, None, None except TypeError: return None, None, None ############## # Properties # ##############
@LazyProperty
[docs] def value(self): return self._load_value(self.force_overwrite) ########### # Methods # ###########
[docs] def flush_value(self): if self.filled: self.value.flush() ################### # Private Methods # ###################
def _load_value(self, force_overwrite=False): load_successful = False if exists(self.filename) and not force_overwrite: try: value = np.load( self.filename, mmap_mode=self.mmap_load_mode ) if self.shape is not None and self.shape != value.shape: raise_warning("Loaded shape {0} does not match specified (or saved) shape {1}".format( value.shape, self.shape ), ShapeMismatchWarning) self.shape = value.shape self.filled = True self.write_mode = False load_successful = True except ValueError as e: raise_warning( "File '{0}' is corrupted and will be overwritten." " Its data will be recomputed. The following error was raised:\n{1}".format( self.filename, traceback.format_exc()), FileOverwriteWarning ) load_successful = False unlink(self.filename) if not load_successful: if self.shape is None: raise ValueError("Can't create cached array if I don't know the shape.") value = open_memmap( self.filename, shape=self.shape, mode='w+', dtype=self.dtype ) self.write_mode = True # Make sure the value is flushed on exit if self.filename not in CachedMemmapArray.files_needing_flush: atexit.register(self.flush_value) CachedMemmapArray.files_needing_flush.add(self.filename) self.filled = False return value
[docs]class CachedComputationUnloader(object): """ Callable, picklable class that CachedComputation.__reduce_ex__ returns an instance of. Pickling and unpickling proceeds by automatic mechanisms """ def __init__(self, comp): """ Argument `pickler_version` is the revision of the file pickling procedure used when the CachedComputation instance was written. """ self.pickler_version = CachedComputation.PICKLE_VERSION self.molecule_xyz = comp.molecule.xyz_string(header=False) self.optional_arguments = comp.optional_arguments.get_simple_dict() self.init_kwargs = dict( basis=comp.basis_name, cached_data=comp.cached_data, psi_options=comp._psi_options, directory=comp.directory ) self.basis_sets_to_register = tuple(comp._basis_registry.keys()) self.allow_analogous_load = comp.allow_analogous_load if hasattr(comp, "minimal_shelf_key"): self.minimal_shelf_key = comp.minimal_shelf_key else: self.minimal_shelf_key = None if hasattr(comp, "shelf_key"): self.shelf_key = comp.shelf_key else: self.shelf_key = None def __call__(self): init_kwargs = dict( # Need to recreate the molecule from the xyz string molecule=self.molecule_xyz, optional_arguments=self.optional_arguments, **self.init_kwargs ) #========================================# rv = CachedComputation(**init_kwargs) #========================================# if self.pickler_version >= (2, 0, 0): for bsname in self.basis_sets_to_register: if self.pickler_version >= (2, 0, 1): # changed _basis_registry from a dict to an AliasedDict, # thus the keys will be frozenset objects bsname = next(iter(bsname)) if bsname not in rv._basis_registry: rv.register_basis(bsname) #----------------------------------------# if self.pickler_version >= (2, 1, 0): if self.minimal_shelf_key is not None: rv.minimal_shelf_key = self.minimal_shelf_key if self.shelf_key is not None: rv.shelf_key = self.shelf_key #----------------------------------------# if self.pickler_version >= (2, 1, 2): rv.allow_analogous_load = self.allow_analogous_load #========================================# return rv #endregion