def _create_newton_system(ode, theta=1):
"""
Return the nonlinear F, the Jacobian matrix and a mapping of states
to nonero fields for the ODE
"""
# Lists of symbols
states = [state.sym for state in ode.states]
states_0 = [state.sym_0 for state in ode.states]
vars_ = [var.sym for var in ode.variables]
vars_0 = [var.sym_0 for var in ode.variables]
def sum_ders(ders):
return reduce(lambda x, y: x + y, ders, 0)
# Substitution dict between sym and previous sym value
subs = [syms for syms in zip(states + vars_, states_0 + vars_0)]
# Generate F using theta rule
F_expr = [
theta * expr + (1 - theta) * expr.subs(subs) -
(sum_ders(ders) - sum_ders(ders).subs(subs)) / ode.dt
for ders, expr in ode.get_derivative_expr(True)
]
# Create mapping of orginal expression
sym_map = OrderedDict()
jac_subs = OrderedDict()
F = []
for i, expr in enumerate(F_expr):
for j, state in enumerate(states):
F_ij = expr.diff(state)
if F_ij:
jac_sym = sp.Symbol(f"j_{i}_{j}")
sym_map[i, j] = jac_sym
jac_subs[jac_sym] = F_ij
# print "[%d,%d] (%d) # [%s, %s] \n%s" \
# % (i,j, len(F_ij.args), states[i], states[j], F_ij)
F_i = sp.Symbol(f"F_{i}")
sym_map[i, j + 1] = F_i
jac_subs[F_i] = expr
F.append(F_i)
# Create the Jacobian
jacobi = sp.SparseMatrix(
len(states),
len(states),
lambda i, j: sym_map.get((i, j), 0),
)
# return Symbolic representation of the linear system
return sp.Matrix(F), F_expr, jacobi, states, jac_subs
def update_entry(new_count, i, j):
global zero_operations
if _iszero(b[j, 0] * A[i, j]):
zero_operations += 1
return new_count
if i != 0:
operations.append(
"dx[{i}] -= dx[{j}]*jac[{i}*{n}+{j}]".format(i=i, j=j,
n=n), )
else:
operations.append("dx[{i}] -= dx[{j}]*jac[{j}]".format(i=i,
j=j))
# Create new symbol and store the representation
new_sym = sp.Symbol(f"F_{i}:{new_count}")
new_old[new_sym] = b[i, 0] - b[j, 0] * A[i, j]
for old_sym in [b[i, 0], b[j, 0], A[i, j]]:
storage = old_new.get(old_sym)
if storage is None:
storage = set()
old_new[old_sym] = storage
storage.add(new_sym)
# Change entry to the new symbol
b[i, 0] = new_sym
new_count += 1
return new_count
def __init__(self, name, unit="ms"):
super(Time, self).__init__(name, ScalarParam(0.0, unit=unit))
# Add previous value symbol
self.sym_0 = sp.Symbol(
f"{name}_0",
real=True,
imaginary=False,
commutative=True,
hermitian=True,
complex=True,
)
def symbol_subs(self):
"""
Return a subs dict for all ODE Objects (states, parameters)
"""
if self._symbol_subs is None:
subs = []
# Deal with parameter subs first
if self.optimization.parameter_numerals:
subs.extend(
(param.sym, param.init) for param in self.ode.parameters)
elif not self.optimization.use_parameter_names:
subs.extend(
(param.sym, sp.Symbol("parameters" + self.index(ind)))
for ind, param in enumerate(self.ode.parameters))
elif self._parameter_prefix:
subs.extend((
param.sym,
sp.Symbol(f"{self._parameter_prefix}{param.name}"),
) for param in self.ode.parameters)
# Deal with state subs
if not self.optimization.use_state_names:
subs.extend((state.sym, sp.Symbol("states" + self.index(ind)))
for ind, state in enumerate(self.ode.states))
elif self._state_prefix:
subs.extend((
param.sym,
sp.Symbol(f"{self._state_prefix}{param.name}"),
) for param in self.ode.states)
self._symbol_subs = subs
return self._symbol_subs
def update_entry(new_count, i, j):
global zero_operations
if _iszero(b[j, 0] * A[i, j]):
zero_operations += 1
return new_count
# Create new symbol and store the representation
new_sym = sp.Symbol(f"F_{i}:{new_count}")
new_old[new_sym] = b[i, 0] - b[j, 0] * A[i, j]
for old_sym in [b[i, 0], b[j, 0], A[i, j]]:
storage = old_new.get(old_sym)
if storage is None:
storage = set()
old_new[old_sym] = storage
storage.add(new_sym)
# Change entry to the new symbol
b[i, 0] = new_sym
new_count += 1
return new_count
def update_entry(new_count, i, j, k):
global zero_operations
if _iszero(A[i, k] * A[k, j]):
zero_operations += 1
return new_count
# Create new symbol and store the representation
new_sym = sp.Symbol(f"j_{i}_{j}:{new_count}")
new_old[new_sym] = A[i, j] - A[i, k] * A[k, j]
for old_sym in [A[i, j], A[i, k], A[k, j]]:
storage = old_new.get(old_sym)
if storage is None:
storage = set()
old_new[old_sym] = storage
storage.add(new_sym)
# Change entry to the new symbol
A[i, j] = new_sym
new_count += 1
return new_count
def _compute_jacobian(self):
if self._jacobian_expr is not None:
return
ode = self.ode
if ode.num_states > 10:
info(
"Calculating jacobian for {0} states. "
"May take some time...".format(ode.num_states), )
sys.stdout.flush()
sym_map = OrderedDict()
jacobi_expr = OrderedDict()
for i, (ders, expr) in enumerate(ode.get_derivative_expr(True)):
for j, state in enumerate(ode.states):
F_ij = expr.diff(state.sym)
# Only collect non zero contributions
if F_ij:
jacobi_sym = sp.Symbol(f"j_{i}_{j}")
sym_map[i, j] = jacobi_sym
jacobi_expr[jacobi_sym] = F_ij
# print "[%d,%d] (%d) # [%s, %s] \n%s" \
# % (i,j, len(F_ij.args), states[i], states[j], F_ij)
# Create the Jacobian
self._jacobian_mat = sp.SparseMatrix(
ode.num_states,
ode.num_states,
lambda i, j: sym_map.get((i, j), 0),
)
if self.optimization.transposed_jacobian:
self._jacobian_mat = self._jacobian_mat.transpose
self._jacobian_expr = jacobi_expr
if ode.num_states > 10:
info(" done")
def __init__(self, name, init, time):
"""
Create a state variable with an associated initial value
Arguments
---------
name : str
The name of the State
init : scalar, ScalarParam
The initial value of this state
time : Time
The time variable
"""
# Call super class
check_arg(init, scalars + (ScalarParam, ), 1, State)
super(State, self).__init__(name, init)
check_arg(time, Time, 2)
self.time = time
# Add previous value symbol
self.sym_0 = sp.Symbol(f"{name}_0")(time.sym)
self.sym_0._assumptions["real"] = True
self.sym_0._assumptions["imaginary"] = False
self.sym_0._assumptions["commutative"] = True
self.sym_0._assumptions["hermitian"] = True
self.sym_0._assumptions["complex"] = True
self._sym = self._param.sym(self.time.sym)
self._sym._assumptions["real"] = True
self._sym._assumptions["imaginary"] = False
self._sym._assumptions["commutative"] = True
self._sym._assumptions["hermitian"] = True
self._sym._assumptions["complex"] = True
# Flag to determine if State is solved or not
self._is_solved = False
def update_entry(new_count, i, j, k):
global zero_operations
if _iszero(A[i, k] * A[k, j]):
zero_operations += 1
return new_count
# Store operation
if i != 0:
operations.append(
"jac[{i}*{n}+{j}] -= jac[{i}*{n}+{k}]*jac[{k}*{n}+{j}]".
format(
i=i,
j=j,
k=k,
n=n,
), )
else:
operations.append(
"jac[{j}] -= jac[{k}]*jac[{k}*{n}+{j}]".format(j=j,
k=k,
n=n), )
new_sym = sp.Symbol(f"j_{i}_{j}:{new_count}")
new_old[new_sym] = A[i, j] - A[i, k] * A[k, j]
for old_sym in [A[i, j], A[i, k], A[k, j]]:
storage = old_new.get(old_sym)
if storage is None:
storage = set()
old_new[old_sym] = storage
storage.add(new_sym)
# Change entry to the new symbol
A[i, j] = new_sym
new_count += 1
return new_count
def _LU_solve(AA, rhs):
"""
Returns the symbolic solve of AA*x=rhs
"""
if not AA.is_square:
raise sympy.NonSquareMatrixError()
n = AA.rows
A = AA[:, :]
p = []
nnz = 0
for i in range(n):
for j in range(n):
nnz += not _iszero(A[i, j])
print("Num non zeros in jacobian:", nnz)
# A map between old symbols and new. The values in this dict corresponds to
# where an old symbol is used. If the length of the value is 1 it is only
# used once and once the value is used it can be freed.
old_new = OrderedDict()
new_old = OrderedDict()
new_count = 0
global zero_operations
zero_operations = 0
def update_entry(new_count, i, j, k):
global zero_operations
if _iszero(A[i, k] * A[k, j]):
zero_operations += 1
return new_count
# Create new symbol and store the representation
new_sym = sp.Symbol(f"j_{i}_{j}:{new_count}")
new_old[new_sym] = A[i, j] - A[i, k] * A[k, j]
for old_sym in [A[i, j], A[i, k], A[k, j]]:
storage = old_new.get(old_sym)
if storage is None:
storage = set()
old_new[old_sym] = storage
storage.add(new_sym)
# Change entry to the new symbol
A[i, j] = new_sym
new_count += 1
return new_count
# factorization
for j in range(n):
for i in range(j):
for k in range(i):
new_count = update_entry(new_count, i, j, k)
pivot = -1
for i in range(j, n):
for k in range(j):
new_count = update_entry(new_count, i, j, k)
# find the first non-zero pivot, includes any expression
if pivot == -1 and not _iszero(A[i, j]):
pivot = i
if pivot < 0:
# this result is based on iszerofunc's analysis of the
# possible pivots, so even though the element may not be
# strictly zero, the supplied iszerofunc's evaluation gave
# True
raise ValueError("No nonzero pivot found; inversion failed.")
if pivot != j: # row must be swapped
A.row_swap(pivot, j)
p.append([pivot, j])
scale = 1 / A[j, j]
for i in range(j + 1, n):
if _iszero(A[i, j]):
zero_operations += 1
continue
# Create new symbol and store the representation
new_sym = sp.Symbol(f"j_{i}_{j}:{new_count}")
new_old[new_sym] = A[i, j] * scale
for old_sym in [A[i, j], A[j, j]]:
storage = old_new.get(old_sym)
if storage is None:
storage = set()
old_new[old_sym] = storage
storage.add(new_sym)
# Change entry to the new symbol
A[i, j] = new_sym
new_count += 1
nnz = 0
for i in range(n):
for j in range(n):
nnz += not _iszero(A[i, j])
print("Num non zeros in factorized jacobian:", nnz)
print("Num non-zero operations while factorizing matrix:", new_count)
print("Num zero operations while factorizing matrix:", zero_operations)
factorizing_nnz_operations = new_count
zero_operations = 0
n = AA.rows
b = rhs.permuteFwd(p)
def update_entry(new_count, i, j):
global zero_operations
if _iszero(b[j, 0] * A[i, j]):
zero_operations += 1
return new_count
# Create new symbol and store the representation
new_sym = sp.Symbol(f"F_{i}:{new_count}")
new_old[new_sym] = b[i, 0] - b[j, 0] * A[i, j]
for old_sym in [b[i, 0], b[j, 0], A[i, j]]:
storage = old_new.get(old_sym)
if storage is None:
storage = set()
old_new[old_sym] = storage
storage.add(new_sym)
# Change entry to the new symbol
b[i, 0] = new_sym
new_count += 1
return new_count
# forward substitution, all diag entries are scaled to 1
for i in range(n):
for j in range(i):
new_count = update_entry(new_count, i, j)
# b.row(i, lambda x,k: x - b[j,k]*A[i,j])
# backward substitution
for i in range(n - 1, -1, -1):
for j in range(i + 1, n):
new_count = update_entry(new_count, i, j)
# b.row(i, lambda x,k: x - b[j,k]*A[i,j])
b.row(i, lambda x, k: x / A[i, i])
print(
"Num zero operations while forward/backward substituting matrix:",
zero_operations,
)
print(
"Num non-zero operations while factorizing matrix:",
new_count - factorizing_nnz_operations,
)
return b, new_old, old_new
return A, p, old_new, new_old
def __init__(
self,
factorized,
function_name="forward_backward_subst",
result_name="dx",
residual_name="F",
params=None,
):
"""
Create a JacobianForwardBackwardSubstComponent
Arguments
---------
factorized : gotran.FactorizedJacobianComponent
The factorized jacobian of the ODE
function_name : str
The name of the function which should be generated
result_name : str
The name of the result (increment)
residual_name : str
The name of the residual
params : dict
Parameters determining how the code should be generated
"""
timer = Timer(
"Computing forward backward substituion component") # noqa: F841
check_arg(factorized, FactorizedJacobianComponent)
jacobian_name = list(factorized.shapes.keys())[0]
descr = ("Symbolically forward backward substitute linear system "
"of {0} ODE".format(factorized.root))
super(ForwardBackwardSubstitutionComponent, self).__init__(
"ForwardBackwardSubst",
factorized.root,
function_name,
descr,
params=params,
use_default_arguments=False,
additional_arguments=[residual_name],
)
self.add_comment(
f"Forward backward substituting factorized linear system {self.root.name}",
)
# Recreate jacobian using only sympy Symbols
jac_orig = factorized.factorized_jacobian
# Size of system
n = jac_orig.rows
jac = sp.Matrix(n, n, lambda i, j: sp.S.Zero)
for i in range(n):
for j in range(n):
# print jac_orig[i,j]
if not jac_orig[i, j].is_zero:
name = sympycode(jac_orig[i, j])
jac[i, j] = sp.Symbol(
name,
real=True,
imaginary=False,
commutative=True,
hermitian=True,
complex=True,
)
print(jac[i, j])
self.shapes[jacobian_name] = (n, n)
self.shapes[residual_name] = (n, )
self.shapes[result_name] = (n, )
F = []
dx = []
# forward substitution, all diag entries are scaled to 1
for i in range(n):
F.append(self.add_indexed_object(residual_name, i))
dx.append(self.add_indexed_expression(result_name, i, F[i]))
for j in range(i):
if jac[i, j].is_zero:
continue
dx[i] = self.add_indexed_expression(
result_name,
i,
dx[i] - dx[j] * jac[i, j],
)
# backward substitution
for i in range(n - 1, -1, -1):
for j in range(i + 1, n):
if jac[i, j].is_zero:
continue
dx[i] = self.add_indexed_expression(
result_name,
i,
dx[i] - dx[j] * jac[i, j],
)
dx[i] = self.add_indexed_expression(result_name, i,
dx[i] / jac[i, i])
# No need to call recreate body expressions
self.body_expressions = [
obj for obj in self.ode_objects
if isinstance(obj, (IndexedExpression, Comment))
]
self.results = [result_name]
self.used_states = set()
self.used_parameters = set()
def __init__(self, jacobian, function_name="lu_factorize", params=None):
"""
Create a FactorizedJacobianComponent
Arguments
---------
jacobian : gotran.JacobianComponent
The Jacobian of the ODE
function_name : str
The name of the function which should be generated
params : dict
Parameters determining how the code should be generated
"""
timer = Timer("Computing factorization of jacobian") # noqa: F841
check_arg(jacobian, JacobianComponent)
descr = f"Symbolically factorize the jacobian of the {jacobian.root} ODE"
super(FactorizedJacobianComponent, self).__init__(
"FactorizedJacobian",
jacobian.root,
function_name,
descr,
params=params,
use_default_arguments=False,
additional_arguments=jacobian.results,
)
self.add_comment(f"Factorizing jacobian of {self.root.name}")
jacobian_name = jacobian.results[0]
# Recreate jacobian using only sympy Symbols
jac_orig = jacobian.jacobian
# Size of system
n = jac_orig.rows
jac = sp.Matrix(n, n, lambda i, j: sp.S.Zero)
for i in range(n):
for j in range(n):
# print jac_orig[i,j]
if not jac_orig[i, j].is_zero:
name = sympycode(jac_orig[i, j])
jac[i, j] = sp.Symbol(
name,
real=True,
imaginary=False,
commutative=True,
hermitian=True,
complex=True,
)
print(jac[i, j])
p = []
self.shapes[jacobian_name] = (n, n)
def add_intermediate_if_changed(jac, jac_ij, i, j):
# If item has changed
if jac_ij != jac[i, j]:
print("jac", i, j, jac_ij)
jac[i,
j] = self.add_indexed_expression(jacobian_name, (i, j),
jac_ij)
# Do the factorization
for j in range(n):
for i in range(j):
# Get sympy expr of A_ij
jac_ij = jac[i, j]
# Build sympy expression
for k in range(i):
jac_ij -= jac[i, k] * jac[k, j]
add_intermediate_if_changed(jac, jac_ij, i, j)
pivot = -1
for i in range(j, n):
# Get sympy expr of A_ij
jac_ij = jac[i, j]
# Build sympy expression
for k in range(j):
jac_ij -= jac[i, k] * jac[k, j]
add_intermediate_if_changed(jac, jac_ij, i, j)
# find the first non-zero pivot, includes any expression
if pivot == -1 and jac[i, j]:
pivot = i
if pivot < 0:
# this result is based on iszerofunc's analysis of the
# possible pivots, so even though the element may not be
# strictly zero, the supplied iszerofunc's evaluation gave
# True
error("No nonzero pivot found; symbolic inversion failed.")
if pivot != j: # row must be swapped
jac.row_swap(pivot, j)
p.append([pivot, j])
print("Pivoting!!")
# Scale with diagonal
if not jac[j, j]:
error("Diagonal element of the jacobian is zero. "
"Inversion failed")
scale = 1 / jac[j, j]
for i in range(j + 1, n):
# Get sympy expr of A_ij
jac_ij = jac[i, j]
jac_ij *= scale
add_intermediate_if_changed(jac, jac_ij, i, j)
# Store factorized jacobian
self.factorized_jacobian = jac
self.num_nonzero = sum(not jac[i, j].is_zero for i in range(n)
for j in range(n))
# No need to call recreate body expressions
self.body_expressions = self.ode_objects
self.used_states = set()
self.used_parameters = set()
def _init_param_state_replace_dict(self):
"""
Create a parameter state replace dict based on the values in the
global parameters
"""
param_state_replace_dict = {}
array_params = self._params["array"]
param_repr = self._params["parameters"]["representation"]
param_name = self._params["parameters"]["array_name"]
param_offset = self._params["parameters"]["add_offset"]
field_param_name = self._params["parameters"]["field_array_name"]
field_param_offset = self._params["parameters"]["add_field_offset"]
field_parameters = self._params["parameters"]["field_parameters"]
# If empty
# FIXME: Get rid of this by introducing a ListParam type in modelparameters
if len(field_parameters) == 1 and field_parameters[0] == "":
field_parameters = []
for param in field_parameters:
if not isinstance(self.root.present_ode_objects[param], Parameter):
error(
f"Field parameter '{param}' is not a parameter in the '{self.root}'",
)
state_repr = self._params["states"]["representation"]
state_name = self._params["states"]["array_name"]
state_offset = self._params["states"]["add_offset"]
time_name = self._params["time"]["name"]
dt_name = self._params["dt"]["name"]
# Create a map between states, parameters and the corresponding
# IndexedObjects
param_state_map = OrderedDict(
[("states", OrderedDict()), ("parameters", OrderedDict())],
)
# Add states
states = param_state_map["states"]
for ind, state in enumerate(self.root.full_states):
states[state] = StateIndexedObject(
state_name,
ind,
state,
(self.root.num_full_states,),
array_params,
state_offset,
)
# Add parameters
parameters = param_state_map["parameters"]
for ind, param in enumerate(self.root.parameters):
if param.name in field_parameters:
basename = field_param_name
index = field_parameters.index(param.name)
shape = (len(field_parameters),)
offset = field_param_offset
else:
basename = param_name
index = ind
shape = (self.num_parameters,)
offset = param_offset
parameters[param] = ParameterIndexedObject(
basename,
index,
param,
shape,
array_params,
offset,
)
# If not having named parameters
if param_repr == "numerals":
param_state_replace_dict.update(
(param.sym, param.init) for param in self.root.parameters
)
elif param_repr == "array":
self.shapes[param_name] = (self.root.num_parameters,)
if field_parameters:
self.shapes["field_" + param_name] = (len(field_parameters),)
param_state_replace_dict.update(
(param.sym, indexed.sym)
for param, indexed in list(param_state_map["parameters"].items())
)
if state_repr == "array":
self.shapes[state_name] = (self.root.num_full_states,)
param_state_replace_dict.update(
(state.sym, indexed.sym)
for state, indexed in list(param_state_map["states"].items())
)
param_state_replace_dict[self.root._time.sym] = sp.Symbol(time_name)
param_state_replace_dict[self.root._dt.sym] = sp.Symbol(dt_name)
self.indexed_map.update(param_state_map)
# return dicts
return param_state_replace_dict
def _compute_symbolic_fb_substitution(self):
if hasattr(self, "_jacobian_fb_substitution_operations"):
return
if not hasattr(self, "_factorized_jacobian"):
self._compute_symbolic_factorization_of_jacobian()
global zero_operations
zero_operations = 0
A = self._factorized_jacobian
n = A.rows
b = sp.Matrix(n, 1, lambda i, j: sp.Symbol(f"F_{i}"))
old_new = OrderedDict()
new_old = OrderedDict()
# b = rhs.permuteFwd(p)
operations = []
new_count = 0
def update_entry(new_count, i, j):
global zero_operations
if _iszero(b[j, 0] * A[i, j]):
zero_operations += 1
return new_count
if i != 0:
operations.append(
"dx[{i}] -= dx[{j}]*jac[{i}*{n}+{j}]".format(i=i, j=j,
n=n), )
else:
operations.append("dx[{i}] -= dx[{j}]*jac[{j}]".format(i=i,
j=j))
# Create new symbol and store the representation
new_sym = sp.Symbol(f"F_{i}:{new_count}")
new_old[new_sym] = b[i, 0] - b[j, 0] * A[i, j]
for old_sym in [b[i, 0], b[j, 0], A[i, j]]:
storage = old_new.get(old_sym)
if storage is None:
storage = set()
old_new[old_sym] = storage
storage.add(new_sym)
# Change entry to the new symbol
b[i, 0] = new_sym
new_count += 1
return new_count
# forward substitution, all diag entries are scaled to 1
for i in range(n):
for j in range(i):
new_count = update_entry(new_count, i, j)
# b.row(i, lambda x,k: x - b[j,k]*A[i,j])
# backward substitution
for i in range(n - 1, -1, -1):
for j in range(i + 1, n):
new_count = update_entry(new_count, i, j)
# b.row(i, lambda x,k: x - b[j,k]*A[i,j])
operations.append("dx[{i}] /= jac[{i}*{n}+{i}]".format(i=i, n=n))
b.row(i, lambda x, k: x / A[i, i])
total_count = zero_operations + new_count * 1.0
if total_count != 0:
print("Num operations while forward/backward substituting matrix:")
print(
"Zero operations:",
zero_operations,
int(100 * zero_operations / total_count),
"%",
)
print(
"Non-zero operations:",
new_count,
int(100 * new_count / total_count),
"%",
)
self._jacobian_fb_substitution_operations = operations
def _compute_symbolic_factorization_of_jacobian(self):
if self._jacobian_expr is None:
self._compute_jacobian_cse()
# Get jacobian
AA = self._jacobian_mat
if not AA.is_square:
raise sympy.NonSquareMatrixError()
# Get copy
n = AA.rows
A = AA[:, :]
p = []
n2 = n * n
nnz = 0
for i in range(n):
# Add dummy value at the diagonal if zero
if _iszero(A[i, i]):
A[i, i] = 1
for j in range(n):
nnz += not _iszero(A[i, j])
print("Num non zeros in jacobian:", nnz, nnz * 1.0 / n2 * 100, "%")
# A map between old symbols and new. The values in this dict corresponds to
# where an old symbol is used. If the length of the value is 1 it is only
# used once and once the value is used it can be freed.
old_new = OrderedDict()
new_old = OrderedDict()
operations = []
new_count = 0
global zero_operations
zero_operations = 0
def update_entry(new_count, i, j, k):
global zero_operations
if _iszero(A[i, k] * A[k, j]):
zero_operations += 1
return new_count
# Store operation
if i != 0:
operations.append(
"jac[{i}*{n}+{j}] -= jac[{i}*{n}+{k}]*jac[{k}*{n}+{j}]".
format(
i=i,
j=j,
k=k,
n=n,
), )
else:
operations.append(
"jac[{j}] -= jac[{k}]*jac[{k}*{n}+{j}]".format(j=j,
k=k,
n=n), )
new_sym = sp.Symbol(f"j_{i}_{j}:{new_count}")
new_old[new_sym] = A[i, j] - A[i, k] * A[k, j]
for old_sym in [A[i, j], A[i, k], A[k, j]]:
storage = old_new.get(old_sym)
if storage is None:
storage = set()
old_new[old_sym] = storage
storage.add(new_sym)
# Change entry to the new symbol
A[i, j] = new_sym
new_count += 1
return new_count
# factorization
for j in range(n):
for i in range(j):
for k in range(i):
new_count = update_entry(new_count, i, j, k)
pivot = -1
for i in range(j, n):
for k in range(j):
new_count = update_entry(new_count, i, j, k)
# find the first non-zero pivot, includes any expression
if pivot == -1 and not _iszero(A[i, j]):
pivot = i
if pivot < 0:
# this result is based on iszerofunc's analysis of the
# possible pivots, so even though the element may not be
# strictly zero, the supplied iszerofunc's evaluation gave
# True
raise ValueError("No nonzero pivot found; inversion failed.")
if pivot != j: # row must be swapped
A.row_swap(pivot, j)
p.append([pivot, j])
print("Pivoting!!")
scale = 1 / A[j, j]
for i in range(j + 1, n):
if _iszero(A[i, j]):
zero_operations += 1
continue
# Store operation
operations.append(
"jac[{i}*{n}+{j}] /= jac[{j}*{n}+{j}]".format(i=i,
n=n,
j=j), )
# Create new symbol and store the representation
new_sym = sp.Symbol(f"j_{i}_{j}:{new_count}")
new_old[new_sym] = A[i, j] * scale
for old_sym in [A[i, j], A[j, j]]:
storage = old_new.get(old_sym)
if storage is None:
storage = set()
old_new[old_sym] = storage
storage.add(new_sym)
# Change entry to the new symbol
A[i, j] = new_sym
new_count += 1
nnz = 0
for i in range(n):
for j in range(n):
nnz += not _iszero(A[i, j])
total_op = new_count + zero_operations
if total_op != 0:
print(
"Num non zeros in factorized jacobian:",
nnz,
int(nnz * 1.0 / n2 * 100),
"%",
)
print(
"Num non-zero operations while factorizing matrix:",
new_count,
new_count * 1.0 / total_op * 100,
"%",
)
print(
"Num zero operations while factorizing matrix:",
zero_operations,
int(zero_operations * 1.0 / total_op * 100),
"%",
)
self._jacobian_factorization_operations = operations
self._factorized_jacobian = A
def __init__(
self,
basename,
indices,
shape=None,
array_params=None,
add_offset="",
dependent=None,
enum=None,
):
"""
Create an IndexedExpression with an associated basename
Arguments
---------
basename : str
The basename of the multi index Expression
indices : tuple of ints
The indices
shape : tuple (optional)
A tuple with the shape of the indexed object
array_params : dict
Parameters to create the array name for the indexed object
add_offset : bool, str
If True a fixed offset is added to the indices
dependent : ODEObject
If given the count of this IndexedObject will follow as a
fractional count based on the count of the dependent object
"""
array_params = array_params or {}
check_arg(basename, str)
indices = tuplewrap(indices)
check_arg(indices, tuple, itemtypes=int)
check_kwarg(array_params, "array_params", dict)
# Get index format and index offset from global parameters
self._array_params = self.default_parameters()
self._array_params.update(array_params)
self._enum = enum
index_format = self._array_params.index_format
index_offset = self._array_params.index_offset
flatten = self._array_params.flatten
if add_offset and isinstance(add_offset, str):
offset_str = add_offset
else:
offset_str = f"{basename}_offset" if add_offset else ""
self._offset_str = offset_str
if offset_str:
offset_str += "+"
# If trying to flatten indices, with a rank larger than 1 a shape needs
# to be provided
if len(indices) > 1 and flatten and shape is None:
error(
"A 'shape' need to be provided to generate flatten indices "
"for index expressions with rank larger than 1.", )
# Create index format
if index_format == "{}":
index_format = "{{{0}}}"
else:
index_format = index_format[0] + "{0}" + index_format[1]
orig_indices = indices
# If flatten indices
if flatten and len(indices) > 1:
indices = (sum(
reduce(lambda i, j: i * j, shape[i + 1:], 1) *
(index + index_offset) for i, index in enumerate(indices)), )
else:
indices = tuple(index + index_offset for index in indices)
index_str = ",".join(offset_str + str(index) for index in indices)
name = basename + index_format.format(index_str)
ODEObject.__init__(self, name, dependent)
self._basename = basename
self._indices = orig_indices
self._sym = sp.Symbol(
name,
real=True,
imaginary=False,
commutative=True,
hermitian=True,
complex=True,
)
self._shape = shape
