twodlearn.linalg module

class twodlearn.linalg.Cholesky(input, name=None)

Bases: twodlearn.core.common.TdlModel

Tdl autoinitialization with arguments:

input

(InputParameter)

linop

(LazzyProperty)

value

(Submodel)

input

linop

log_abs_determinant()

matmul(x, adjoint=False, adjoint_arg=False, name='matmul')

matvec(x, adjoint=False, name='matvec')

shape_tensor()

Shape determined at runtime

solve(rhs, adjoint=False, adjoint_arg=False, name='solve')

solvevec(rhs, adjoint=False, name='solve')

value

class twodlearn.linalg.DynamicScaledIdentity(**kargs)

Bases: twodlearn.core.common.TdlModel

scaled identity [batched] matrix whose size depends on the shape of the input

Tdl autoinitialization with arguments:

tolerance

(InputArgument)

multipliers

(InputParameter) tensor whose elements represent the multiplier factors for the identity matrices.

property batch_shape

multipliers

tensor whose elements represent the multiplier factors for the identity matrices.

tolerance

twodlearn.linalg.M_times_Mt(M)

Compute M @ M^t

Parameters

M – (batched) matrix M

Returns

solution of M @ M^t

Return type

tf.Tensor

twodlearn.linalg.Mt_times_M(M)

Compute M^t @ M

Parameters

M – (batched) matrix M

Returns

solution of M^t @ M

Return type

tf.Tensor

class twodlearn.linalg.PDMatrix(**kargs)

Bases: twodlearn.core.common.TdlModel

Positive definite matrix

Tdl autoinitialization with arguments:

linop

(LazzyProperty)

shape

(InputArgument) shape of the distribution. The shape assumes the last two dimentions corresponds to the shape of a square matrix

cholesky

(LazzyProperty)

domain_dimension

(InputArgument)

raw

(SimpleParameter)

tolerance

(InputArgument)

value

(OutputValue)

batch_shape

(InputArgument)

batch_shape

cholesky

domain_dimension

linop

raw

shape

shape of the distribution. The shape assumes the last two dimentions corresponds to the shape of a square matrix

tolerance

value

class twodlearn.linalg.PDMatrixDiag(**kargs)

Bases: twodlearn.linalg.PDMatrix

Tdl autoinitialization with arguments:

linop

(LazzyProperty)

shape

(InputArgument) shape of the distribution. The shape assumes the last two dimentions corresponds to the shape of a square matrix

cholesky

(LazzyProperty)

domain_dimension

(InputArgument)

raw

(SimpleParameter) square root of the diagonal part for the matrices

tolerance

(InputArgument)

value

(OutputValue)

batch_shape

(InputArgument)

linop

raw

square root of the diagonal part for the matrices

class twodlearn.linalg.PDScaledIdentity(**kargs)

Bases: twodlearn.linalg.PDMatrixDiag

Tdl autoinitialization with arguments:

raw

(InputParameter) multiplier for the identities

shape

(InputArgument) shape of the distribution. The shape assumes the last two dimentions corresponds to the shape of a square matrix

linop

(SubmodelInit) Autoinit with arguments [‘domain_dimension’]

cholesky

(LazzyProperty)

domain_dimension

(InputArgument)

tolerance

(InputArgument)

value

(OutputValue)

batch_shape

(InputArgument)

domain_dimension

linop

Autoinit with arguments [‘domain_dimension’]

raw

multiplier for the identities

twodlearn.linalg.diagonal_M_times_Mt(M)

Compute M @ M^t for diagonal linop M.

Parameters

M – (batched) diagonal linear operator

Returns

solution of M @ M^t

Return type

tf.Tensor

twodlearn.linalg.is_diagonal(x)

twodlearn.linalg.is_diagonal_linop(M)

twodlearn.linalg.is_scaled_identity(x)

twodlearn.linalg.solvemat(M_cholesky, A)

Compute inv(M) @ A

Parameters

• M_cholesky – cholesky decomposition of matrix M

• A (tf.Tensor) – (batched) Matrix.

Returns

solution of inv(M) @ A

Return type

tf.Tensor

twodlearn.linalg.solvevec(M_cholesky, x)

compute inv(M) @ x

Parameters

• M_cholesky – cholesky decomposition of matrix M.

• x (tf.Tensor) – (batched) vector

Returns

solution of inv(M) @ x

Return type

tf.Tensor