def LDL(A):
if not isinstance(A, Dense):
raise Exception("Matrix A must be dense")
A = copy.deepcopy(A)
L = Dense(A.rows, A.columns)
D = Vector(A.rows)
for i in range(A.rows):
L.data[i][i] = 1
a = A.data[i][i]
for j in range(i):
a = a - D.data[j] * L.data[i][j] * L.data[i][j]
D.data[i] = a
if D.data[i] == 0:
raise Exception("A is no LDL decomposition")
for j in range(i + 1, A.rows):
a = A.data[j][i]
for k in range(j):
a = a - D.data[k] * L.data[j][k] * L.data[i][k]
L.data[j][i] = a / D.data[i]
return [L, D]
def test_vector_setup():
from matrix import Vector
v_1 = Vector(3)
v_1[0] = 5
v_1[2] = -3
assert str(v_1) == '[5, 0, -3]'
v_2 = Vector(3, value=[2, 1, 3])
assert v_2[0] == 2 and v_2[1] == 1 and v_2[2] == 3
with pytest.raises(Exception) as _:
_ = Vector(4, value=[2, 1, 3])
def test_vector_add():
from matrix import Vector
v_1 = Vector(3)
v_1[0] = 5
v_1[2] = -3
v_bad = Vector(4, value=[2, 1, 3, 4])
with pytest.raises(Exception) as _:
_ = Vector.add(v_1, v_bad)
v_2 = Vector(3, value=[-5, 0, 3])
v_add = Vector.add(v_1, v_2)
assert v_add[0] == 0 and v_add[1] == 0 and v_add[2] == 0
def test_vector_dot():
from matrix import Vector
v_1 = Vector(3)
v_1[0] = 5
v_1[2] = -3
v_bad = Vector(4, value=[2, 1, 3, 4])
with pytest.raises(Exception) as _:
_ = Vector.dot(v_1, v_bad)
v_2 = Vector(3, value=[-5, 0, 3])
v_dot = Vector.dot(v_1, v_2)
assert v_dot == -34
def fromAdjacencyToEdge(A, rand):
duplicates = set()
entries = []
w = []
for i in range(A.rows):
for k in range(A.rowPtr[i], A.rowPtr[i + 1]):
j = A.colInd[k]
entry = (i, j)
#BUG something breaks when [1] -> entry[1]
if entry not in duplicates and entry[::
-1] not in duplicates and entry[
0] != entry[1]:
duplicates.add(entry)
entries.append(sorted(entry))
if rand:
w.append(random.random())
else:
w.append(A.data[k])
entries = sorted(entries, key=itemgetter(0))
return [
_processEntries(len(entries), A.columns, entries),
Vector(len(w), w)
]
def factorization(a: Matrix) -> (Matrix, Matrix):
"""Get Q and R from QR factorization"""
size, size = a.size()
q = Matrix.identity(size)
r = a.copy()
for i in range(size - 1):
a_i = Vector(size)
col = r.column(i)
for j in range(i, r.num_columns()):
a_i[j] = col[j]
delta = a_i[i] / abs(a_i[i])
v_i = a_i + delta * a_i.norm() * canonical_base(size, i)
h_v = householder(v_i)
q = q * h_v
r = h_v * r
return q, r
def constructComponent(A, components=[]):
if len(components) == 0:
components = [l1Solver(A)]
b = Vector(A.columns)
xInit = Vector(A.columns)
x = Vector.fromRandomn(A.columns)
compositor = composite(A, xInit, components)
for i in range(10):
x = x + compositor(b)
w = x.scale(1 / x.norm())
component = _formComponent(A, w)
return component, w
def delta_x(x):
mat = Matrix.from_list(jacobi(x))
B = Vector.from_list([-f1(x), -f2(x)])
LU, P = lu.LU_decomposition(mat)
new_B = lu.get_new_B(B, P)
delta = lu.LU_solve(LU, new_B)
return delta.get_data()
def l1Smoother(A):
if not isinstance(A, Sparse):
raise Exception("Matrix A must be sparse")
if A.rows != A.columns:
raise Exception("Matrix A must be square and SPD")
b = Vector(A.rows)
for i in range(A.rows):
sum = 0
for k in range(A.rowPtr[i], A.rowPtr[i + 1]):
sum += abs(A.data[k])
b.data[i] = sum
return b
def diagonal(A):
if not isinstance(A, Sparse):
raise Exception("Matrix A must be sparse")
if A.rows != A.columns:
raise Exception("Matrix A must be square and SPD")
D = Vector(A.rows)
for i in range(A.rows):
for k in range(A.rowPtr[i], A.rowPtr[i + 1]):
j = A.colInd[k]
if i == j:
D.data[i] = A.data[k]
return D
def tc_Series_Vector(source_object, item='data'):
'''
Function to convert from dataframe.Series object to a matrix.Vector
object.
@param source_object: object to be type casted / converted.
@type source_object: dataframe.Series object
@param item: item type to be converted into a list. Allowable items are
'data' and 'label'. Default = data.
@type item: string
@return: matrix.Vector object.
'''
if item == 'data':
return Vector(source_object.data)
if item == 'label':
return Vector(source_object.label)
def gradient(self):
t = []
for i in range(self.weights.size[0]):
t += [
self.weights[i] -
(self.learning_rate * 1 / self.m * self.gradient_error(i))
]
self.weights = Vector(t)
def __init__(self, x, y, cost=1, learning_rate=0.2):
self.x = Matrix([[1, *sample] for sample in x.raw])
self.y = y
self.m = self.x.size[0]
self.n = self.x.size[1]
self.weights = Vector.gen(self.n, fill=0)
self.learning_rate = learning_rate
self.stop_cost = cost
def LU_solve(LU, B):
sz = len(LU)
z = Vector(sz)
x = Vector(sz)
z[0] = B[0]
for i in range(1, sz):
s = sum([LU[i][j] * z[j] for j in range(i)])
z[i] = B[i] - s
x[-1] = z[-1] / LU[-1][-1]
for i in reversed(range(sz - 1)):
s = sum([LU[i][j] * x[j]
for j in range(i + 1, sz)])
x[i] = (z[i] - s) / LU[i][i]
return x
def iterative_method(alpha, beta, eps):
x = Vector.copy(beta)
norm_alpha = alpha.norm()
while True:
x_i = beta + alpha.multiply(x)
if finish_iter_process(x_i, x, norm_alpha, norm_alpha, eps):
break
x = x_i
return x_i
def _solveCoarse(A, r):
converted = csr_matrix((A.data, A.colInd, A.rowPtr),
(A.rows, A.columns)).asfptype()
B = splu(converted.tocsc())
y = B.solve(np.array(r.data))
y = Vector(y.shape[0], list(y))
return y
def hw5(filename, O, maxIter, tolerance, preconditioner, M, display,
displayResidual):
with open(util.getMatrixFile(filename)) as file:
A = Sparse.fromFile(file, O)
if preconditioner.lower() == "d":
method = "Diagonal"
preconditioner = diagonalPreconditioner(A)
elif preconditioner.lower() == "sgs":
method = "Symmetric Gauss Seidel"
preconditioner = sgsPreconditioner(A)
elif preconditioner.lower() == "2l":
method = "Symmetric Two Level"
if M == "l1":
M = l1Solver(A)
method += " (L1 Smoother)"
elif M == "fgs":
M = fgsSolver(A)
method += " (Forward Gauss Seidel)"
preconditioner = twolevelPreconditioner(A, M)
else:
raise Exception("No valid preconditioner selected")
# Generate random solution vector
x = Vector.fromRandom(A.columns, 0, 5)
b = A.multVec(x)
# Generate first iteration
xInit = Vector(A.columns)
start = time.time()
xResult, iterations, residual = pcg(A, b, xInit, maxIter, tolerance,
preconditioner, displayResidual)
end = time.time()
convergence = "(Convergence)" if maxIter != iterations else ""
if display:
util.compareVectors(x, xResult)
print(f"\nPreconditioner = {method}")
print(f"Iterations = {iterations} {convergence}")
print(f"Residual = {residual}")
print(f"Time to solve was {end - start}\n")
def getVertexAggregate(X, clusters):
data = []
colInd = []
rowPtr = [0]
nnz = 0
X = [Vector(len(x), list(x)) for x in X]
for i in range(len(clusters)):
clusters[i] = [Vector(len(c), list(c)) for c in clusters[i]]
for x in X:
aggregate = _findAggregate(x, clusters)
data.append(1)
colInd.append(aggregate)
nnz += 1
rowPtr.append(nnz)
return Sparse(nnz, len(clusters), data, colInd, rowPtr)
def predict(self, sample, poly=False):
if sample.size[1] == self.weights.size[0]:
if poly == True:
prediction = (Vector([f**i for i, f in enumerate(sample.raw)],
transpose=True) * self.weights)[0][0]
else:
prediction = (sample * self.weights)[0][0]
return prediction
else:
print('Wrong Sample', sample)
def hw2(fileName, O, maxIter, tolerance, iterMatrix, display, displayResidual):
with open(util.getMatrixFile(fileName)) as file:
A = Sparse.fromFile(file, O)
# Generate random solution vector
x = Vector.fromRandom(A.columns, 0, 5)
b = A.multVec(x)
# Generate first iteration
xInit = Vector(A.columns)
# Get proper iteration matrix
if iterMatrix.lower() == "l1":
method = "l1 Smoother"
iterSolver = l1Solver(A)
elif iterMatrix.lower() == "fgs":
method = "Forward Gauss Seidel"
iterSolver = fgsSolver(A)
elif iterMatrix.lower() == "bgs":
method = "Backward Gauss Seidel"
iterSolver = bgsSolver(A)
elif iterMatrix.lower() == "sgs":
method = "Symmetric Gauss Seidel"
iterSolver = sgsSolver(A)
else:
raise Exception("No valid iteration matrix selected")
# Solve system using stationary iterative method
start = time.time()
xResult, iterations, residual, accuracy = stationaryIterative(
A, b, xInit, maxIter, tolerance, iterSolver, displayResidual)
end = time.time()
convergence = "(Convergence)" if maxIter != iterations else ""
if display:
util.compareVectors(x, xResult)
print(f"\nB Matrix = {method}")
print(f"Iterations = {iterations} {convergence}")
print(f"Residual = {residual}")
print(f"Accuracy = {accuracy}")
print(f"Time to solve was {end - start}\n")
def get_c(x, f, h):
n = len(f)
a = [0] + [h[i - 1] for i in range(3, n)]
b = [2 * (h[i - 1] + h[i]) for i in range(2, n)]
c = [h[i] for i in range(2, n - 1)] + [0]
d = [
3 * ((f[i] - f[i - 1]) / h[i] - ((f[i - 1] - f[i - 2]) / h[i - 1]))
for i in range(2, n)
]
x = tma(TridiagonalMatrix.from_lists(a, b, c), Vector.from_list(d))
res = [0, 0] + x.get_data()
return res
def learn(self, iters=10000, s_cost=0.1, d_cost=0.01, polynomial=False):
cost = []
for i in range(iters + 1):
# calculating iteration predictions
self.predictions = Vector(
[self.predict(self.x[i], polynomial) for i in range(self.m)])
# applying gradient descent
self.gradient()
# calculating new cost
cost += [self.error()]
print("\rIteration : {}, Cost: {}".format(i, cost[len(cost) - 1]),
end="")
if cost[len(cost) - 1] < s_cost:
break
if len(cost) > 50 and (cost[len(cost) - 50] -
cost[len(cost) - 1]) < d_cost:
break
print('\n')
self.predictions = Vector(
[self.predict(self.x[i], polynomial) for i in range(self.m)])
self.cost = Vector(cost)
def vectorSolver(a, b):
if a.dim != b.dim:
raise Exception(f"Dimension mismatch, {a.dim} != {b.dim}")
data = [0] * a.dim
for i in range(a.dim):
if a.data[i] == 0 or b.data[i] == 0:
data[i] = 0
else:
data[i] = b.data[i] / a.data[i]
return Vector(a.dim, data)
def finite_difference_method(f, p, q, a, b, alpha, beta, delta, gamma, y0, y1,
h):
n = int((b - a) / h)
x = [i for i in np.arange(a, b + h, h)]
A = [0] + [1 - p(x[i]) * h / 2 for i in range(0, n - 1)] + [-gamma]
B = [alpha * h - beta] + [q(x[i]) * h**2 - 2
for i in range(0, n - 1)] + [delta * h + gamma]
C = [beta] + [1 + p(x[i]) * h / 2 for i in range(0, n - 1)] + [0]
D = [y0 * h] + [f(x[i]) * h**2 for i in range(0, n - 1)] + [y1 * h]
y = tma(TridiagonalMatrix.from_lists(A, B, C), Vector.from_list(D))
return x, y.get_data()
def equivalent_form(A, B):
sz = len(A)
alpha = Matrix.zero(sz)
beta = Vector(sz)
# чтобы не делить на 0, добавим небольшую константу
for i in range(sz):
beta[i] = B[i] / (A[i][i] + 1e-3)
for j in range(sz):
if i != j:
alpha[i][j] = -A[i][j] / (A[i][i] + 1e-3)
return alpha, beta
def mls(n, x, y):
N = len(x)
mat = [[sum([x_j**(i + j) for x_j in x]) for i in range(n + 1)]
for j in range(n + 1)]
mat[0][0] = N + 1
b = [sum([x_j**i * y_j for x_j, y_j in zip(x, y)]) for i in range(n + 1)]
mat = Matrix.from_list(mat)
B = Vector.from_list(b)
LU, P = lu.LU_decomposition(mat)
new_B = lu.get_new_B(B, P)
coefs = lu.LU_solve(LU, new_B)
return coefs.get_data()
def seidel_method(alpha, beta, eps):
sz = len(alpha)
np_alpha = np.array(alpha.get_data())
np_beta = np.array(beta.get_data())
B = np.tril(np_alpha, -1)
C = np_alpha - B
tmp1 = inv(np.eye(sz, sz) - B) @ C
tmp2 = inv(np.eye(sz, sz) - B) @ np_beta
x = tmp2
while True:
x_i = tmp2 + tmp1 @ x
if finish_iter_process(x_i, x, norm(tmp1), norm(C), eps, zeidel=True):
break
x = x_i
return Vector.from_list(x_i.tolist())
def normalize(self, data):
if type(data).__name__ == 'Matrix':
data_norm = []
for c in range(data.size[1]):
column = data(c)
min_ = min(column.raw)
max_ = max(column.raw)
data_norm += [[(data[l][c] - min_) / (max_ - min_)
for l in range(data.size[0])]]
return Matrix(data_norm).transpose()
elif type(data).__name__ == 'Vector':
min_ = min(data.raw)
max_ = max(data.raw)
return Vector([(data[l] - min_) / (max_ - min_)
for l in range(data.size[0])])
def hw1(fileName, O, display):
with open(util.getMatrixFile(fileName)) as file:
A = Dense.fromFile(file, O)
# Convert matrix to np array so I can use scipy factorization
A = np.array(A.data)
# Get LDLt factorization
start = time.time()
L, D, P = linalg.ldl(A)
U = np.transpose(L)
end = time.time()
print(f"\nTime to factor input matrix was {end - start} seconds")
# Convert to CSR format used in my library
L = Sparse.fromDense(Dense(L.shape[0], L.shape[1], L.tolist()))
U = Sparse.fromDense(Dense(U.shape[0], U.shape[1], U.tolist()))
# Generate random solution vector
x = Vector.fromRandom(L.columns, 0, 5)
# Resulting b matrix times x vector
bL = L.multVec(x)
bU = U.multVec(x)
# Solve lower triangular system
start = time.time()
r1 = forwardSparse(L, bL)
end = time.time()
print(f"Time to solve lower system was {end - start} seconds")
# Solve upper triangular system
start = time.time()
r2 = backwardSparse(U, bU)
end = time.time()
print(f"Time to solve upper system was {end - start} seconds\n")
if display:
print("Solution to lower triangular")
util.compareVectors(x, r1)
print("\nSolution to upper triangular")
util.compareVectors(x, r2)
def tma(mat, D):
sz = len(mat)
x = Vector(sz)
p, q = [], []
p.append(-mat.c[0] / mat.b[0])
q.append(D[0] / mat.b[0])
for i in range(1, sz):
p_i = 0 if i == sz - 1 else (-mat.c[i] / (mat.b[i] + mat.a[i] * p[i - 1]))
q_i = (D[i] - mat.a[i] * q[i - 1]) / (mat.b[i] + mat.a[i] * p[i - 1])
p.append(p_i)
q.append(q_i)
x[sz - 1] = q[sz - 1]
for i in range(sz - 2, -1, -1):
x[i] = p[i] * x[i + 1] + q[i]
return x
