Skip to content

P
Paddle

PaddlePaddle / Paddle
1 年多 前同步成功

通知 2302
Star 20931
Fork 5422
P
Paddle
未验证 提交 632a0064 编写于 作者: Z zhangkaihuo 提交者： GitHub
浏览文件
操作

[cherry-pick] update multi_dot exposure rules (#36018) (#36131) 

根据线性代数库的API暴露规则修改multi_dot的API暴露规则：
1、在python/paddle/tensor/linalg.py 路径下实现
2、在python/paddle/linalg.py 下import并加入__all__列表
3、在python/paddle/tensor/init.py下引入并加入tensor_method_func列表
4、删除了pythonpaddle/init.py的import
上级 c576169b
隐藏空白更改
内联 并排
Showing with 68 addition and 67 deletion
python/paddle/tensor/__init__.py
浏览文件 @ 632a0064
......@@ -387,6 +387,7 @@ tensor_method_func = [ #noqa
'bitwise_not',
'broadcast_tensors',
'uniform_',
'multi_dot',
'solve',
]
......
python/paddle/tensor/linalg.py
浏览文件 @ 632a0064
......@@ -551,8 +551,8 @@ def cond(x, p=None, name=None):
Computes the condition number of a matrix or batches of matrices with respect to a matrix norm ``p``.
Args:
x (Tensor): The input tensor could be tensor of shape ``(*, m, n)`` where ``*`` is zero or more batch dimensions
for ``p`` in ``(2, -2)``, or of shape ``(*, n, n)`` where every matrix is invertible for any supported ``p``.
x (Tensor): The input tensor could be tensor of shape ``(*, m, n)`` where ``*`` is zero or more batch dimensions
for ``p`` in ``(2, -2)``, or of shape ``(*, n, n)`` where every matrix is invertible for any supported ``p``.
And the input data type could be ``float32`` or ``float64``.
p (float|string, optional): Order of the norm. Supported values are `fro`, `nuc`, `1`, `-1`, `2`, `-2`,
`inf`, `-inf`. Default value is `None`, meaning that the order of the norm is `2`.
......@@ -607,7 +607,7 @@ def cond(x, p=None, name=None):
# out_minus_inf.numpy() [1.]
a = paddle.to_tensor(np.random.randn(2, 4, 4).astype('float32'))
# a.numpy()
# a.numpy()
# [[[ 0.14063153 -0.996288 0.7996131 -0.02571543]
# [-0.16303636 1.5534962 -0.49919784 -0.04402903]
# [-1.1341571 -0.6022629 0.5445269 0.29154757]
......@@ -975,8 +975,8 @@ def t(input, name=None):
return out
check_variable_and_dtype(
input, 'input', ['float16', 'float32', 'float64', 'int32', 'int64'],
'transpose')
input, 'input', ['float16', 'float32', 'float64', 'int32',
'int64'], 'transpose')
helper = LayerHelper('t', **locals())
out = helper.create_variable_for_type_inference(input.dtype)
......@@ -1108,17 +1108,17 @@ def matrix_rank(x, tol=None, hermitian=False, name=None):
r"""
Computes the rank of a matrix.
The rank of a matrix is the number of singular values that are greater than the specified `tol` threshold when hermitian=False,
The rank of a matrix is the number of singular values that are greater than the specified `tol` threshold when hermitian=False,
or the number of eigenvalues in absolute value that are greater than the specified `tol` threshold when hermitian=True.
Args:
x (Tensor): The input tensor. Its shape should be `[..., m, n]`, where `...` is zero or more batch dimensions. If `x` is a batch
of matrices then the output has the same batch dimensions. The data type of `x` should be float32 or float64.
tol (float,Tensor,optional): the tolerance value. Default: None. If `tol` is not specified, and `sigma` is the largest
singular value (or eigenvalues in absolute value), and `eps` is the epsilon value for the dtype of `x`, then `tol` is computed
x (Tensor): The input tensor. Its shape should be `[..., m, n]`, where `...` is zero or more batch dimensions. If `x` is a batch
of matrices then the output has the same batch dimensions. The data type of `x` should be float32 or float64.
tol (float,Tensor,optional): the tolerance value. Default: None. If `tol` is not specified, and `sigma` is the largest
singular value (or eigenvalues in absolute value), and `eps` is the epsilon value for the dtype of `x`, then `tol` is computed
with formula `tol=sigma * max(m,n) * eps`. Note that if `x` is a batch of matrices, `tol` is computed this way for every batch.
hermitian (bool,optional): indicates whether `x` is Hermitian. Default: False. When hermitian=True, `x` is assumed to be Hermitian,
enabling a more efficient method for finding eigenvalues, but `x` is not checked inside the function. Instead, We just use
hermitian (bool,optional): indicates whether `x` is Hermitian. Default: False. When hermitian=True, `x` is assumed to be Hermitian,
enabling a more efficient method for finding eigenvalues, but `x` is not checked inside the function. Instead, We just use
the lower triangular of the matrix to compute.
name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.
......@@ -1225,7 +1225,7 @@ def bmm(x, y, name=None):
#output value:
#[[[6.0, 6.0],[12.0, 12.0]],[[45.0, 45.0],[60.0, 60.0]]]
out_np = out.numpy()
"""
x_shape = x.shape
y_shape = y.shape
......@@ -1360,7 +1360,7 @@ def det(x, name=None):
Returns:
y (Tensor):the determinant value of a square matrix or batches of square matrices.
Example:
Examples:
.. code-block:: python
import paddle
......@@ -1370,10 +1370,10 @@ def det(x, name=None):
A = paddle.linalg.det(x)
print(A)
# [ 0.02547996, 2.52317095, -6.15900707])
"""
if in_dygraph_mode():
return core.ops.determinant(x)
......@@ -1403,7 +1403,7 @@ def slogdet(x, name=None):
"""
Calculates the sign and natural logarithm of the absolute value of a square matrix's or batches square matrices' determinant.
The determinant can be computed with ``sign * exp(logabsdet)
Supports input of float, double
Note that for matrices that have zero determinant, this returns ``(0, -inf)``
......@@ -1415,7 +1415,7 @@ def slogdet(x, name=None):
y (Tensor): A tensor containing the sign of the determinant and the natural logarithm
of the absolute value of determinant, respectively.
Example:
Examples:
.. code-block:: python
import paddle
......@@ -1425,7 +1425,7 @@ def slogdet(x, name=None):
A = paddle.linalg.slogdet(x)
print(A)
# [[ 1. , 1. , -1. ],
# [-0.98610914, -0.43010661, -0.10872950]])
......@@ -1461,19 +1461,19 @@ def svd(x, full_matrices=False, name=None):
Let :math:`X` be the input matrix or a batch of input matrices, the output should satisfies:
.. math::
X = U * diag(S) * VT
X = U * diag(S) * VT
Args:
x (Tensor): The input tensor. Its shape should be `[..., N, M]`,
where `...` is zero or more batch dimensions. N and M can be arbitraty
positive number. Note that if x is sigular matrices, the grad is numerical
instable. The data type of x should be float32 or float64.
full_matrices (bool): A flag to control the behavor of svd.
If full_matrices = True, svd op will compute full U and V matrics,
positive number. Note that if x is sigular matrices, the grad is numerical
instable. The data type of x should be float32 or float64.
full_matrices (bool): A flag to control the behavor of svd.
If full_matrices = True, svd op will compute full U and V matrics,
which means shape of U is `[..., N, N]`, shape of V is `[..., M, M]`. K = min(M, N).
If full_matrices = False, svd op will use a economic method to store U and V.
If full_matrices = False, svd op will use a economic method to store U and V.
which means shape of U is `[..., N, K]`, shape of V is `[..., M, K]`. K = min(M, N).
name (str, optional): Name for the operation (optional, default is None).
name (str, optional): Name for the operation (optional, default is None).
For more information, please refer to :ref:`api_guide_Name`.
Returns:
......@@ -1497,9 +1497,9 @@ def svd(x, full_matrices=False, name=None):
print (vh)
#VT= [[ 0.51411221, 0.85772294],
# [ 0.85772294, -0.51411221]]
# one can verify : U * S * VT == X
# U * UH == I
# U * UH == I
# V * VH == I
"""
......@@ -1526,7 +1526,7 @@ def svd(x, full_matrices=False, name=None):
def matrix_power(x, n, name=None):
r"""
Computes the n-th power of a square matrix or a batch of square matrices.
Let :math:`X` be a sqaure matrix or a batch of square matrices, :math:`n` be
an exponent, the equation should be:
......@@ -1596,27 +1596,27 @@ def matrix_power(x, n, name=None):
def eigvals(x, name=None):
"""
Compute the eigenvalues of one or more general matrices.
Warning:
The gradient kernel of this operator does not yet developed.
Warning:
The gradient kernel of this operator does not yet developed.
If you need back propagation through this operator, please replace it with paddle.linalg.eig.
Args:
x (Tensor): A square matrix or a batch of square matrices whose eigenvalues will be computed.
Its shape should be `[*, M, M]`, where `*` is zero or more batch dimensions.
Its shape should be `[*, M, M]`, where `*` is zero or more batch dimensions.
Its data type should be float32, float64, complex64, or complex128.
name (str, optional): Name for the operation (optional, default is None).
name (str, optional): Name for the operation (optional, default is None).
For more information, please refer to :ref:`api_guide_Name`.
Returns:
Tensor: A tensor containing the unsorted eigenvalues which has the same batch dimensions with `x`.
Tensor: A tensor containing the unsorted eigenvalues which has the same batch dimensions with `x`.
The eigenvalues are complex-valued even when `x` is real.
Examples:
.. code-block:: python
import paddle
paddle.set_device("cpu")
paddle.seed(1234)
......@@ -1630,8 +1630,8 @@ def eigvals(x, name=None):
"""
check_variable_and_dtype(x, 'dtype',
['float32', 'float64', 'complex64', 'complex128'],
'eigvals')
['float32', 'float64', 'complex64',
'complex128'], 'eigvals')
x_shape = list(x.shape)
if len(x_shape) < 2:
......@@ -1657,7 +1657,7 @@ def multi_dot(x, name=None):
"""
Multi_dot is an operator that calculates multiple matrix multiplications.
Supports inputs of float, double and float16 dtypes. This function does not
Supports inputs of float16(only GPU support), float32 and float64 dtypes. This function does not
support batched inputs.
The input tensor in [x] must be 2-D except for the first and last can be 1-D.
......@@ -1735,7 +1735,7 @@ def multi_dot(x, name=None):
def eigh(x, UPLO='L', name=None):
"""
Compute the eigenvalues and eigenvectors of a
Compute the eigenvalues and eigenvectors of a
complex Hermitian (conjugate symmetric) or a real symmetric matrix.
Args:
......@@ -1804,7 +1804,7 @@ def eigh(x, UPLO='L', name=None):
def pinv(x, rcond=1e-15, hermitian=False, name=None):
r"""
Calculate pseudo inverse via SVD(singular value decomposition)
Calculate pseudo inverse via SVD(singular value decomposition)
of one matrix or batches of regular matrix.
.. math::
......@@ -1815,30 +1815,30 @@ def pinv(x, rcond=1e-15, hermitian=False, name=None):
else:
x = u * s * ut (eigh)
out = u * 1/s * u.conj().transpose(-2,-1)
If x is hermitian or symmetric matrix, svd will be replaced with eigh.
Args:
x(Tensor): The input tensor. Its shape should be (*, m, n)
where * is zero or more batch dimensions. m and n can be
arbitraty positive number. The data type of x should be
x(Tensor): The input tensor. Its shape should be (*, m, n)
where * is zero or more batch dimensions. m and n can be
arbitraty positive number. The data type of x should be
float32 or float64 or complex64 or complex128. When data
type is complex64 or cpmplex128, hermitian should be set
True.
rcond(Tensor, optional): the tolerance value to determine
when is a singular value zero. Defalut:1e-15.
hermitian(bool, optional): indicates whether x is Hermitian
rcond(Tensor, optional): the tolerance value to determine
when is a singular value zero. Defalut:1e-15.
hermitian(bool, optional): indicates whether x is Hermitian
if complex or symmetric if real. Default: False.
name(str|None): A name for this layer(optional). If set None,
name(str|None): A name for this layer(optional). If set None,
the layer will be named automatically.
Returns:
Tensor: The tensor with same data type with x. it represents
Tensor: The tensor with same data type with x. it represents
pseudo inverse of x. Its shape should be (*, n, m).
Examples:
.. code-block:: python
......@@ -1998,8 +1998,8 @@ def pinv(x, rcond=1e-15, hermitian=False, name=None):
helper = LayerHelper('pinv', **locals())
dtype = x.dtype
check_variable_and_dtype(
x, 'dtype', ['float32', 'float64', 'complex64', 'complex128'],
'pinv')
x, 'dtype', ['float32', 'float64', 'complex64',
'complex128'], 'pinv')
if dtype == paddle.complex128:
s_type = 'float64'
......@@ -2079,40 +2079,40 @@ def solve(x, y, name=None):
Computes the solution of a square system of linear equations with a unique solution for input 'X' and 'Y'.
Let :math: `X` be a sqaure matrix or a batch of square matrices, :math:`Y` be
a vector/matrix or a batch of vectors/matrices, the equation should be:
.. math::
Out = X^-1 * Y
Specifically,
- This system of linear equations has one solution if and only if input 'X' is invertible.
Args:
x (Tensor): A square matrix or a batch of square matrices. Its shape should be `[*, M, M]`, where `*` is zero or
more batch dimensions. Its data type should be float32 or float64.
y (Tensor): A vector/matrix or a batch of vectors/matrices. Its shape should be `[*, M, K]`, where `*` is zero or
more batch dimensions. Its data type should be float32 or float64.
name(str, optional): Name for the operation (optional, default is None).
name(str, optional): Name for the operation (optional, default is None).
For more information, please refer to :ref:`api_guide_Name`.
Returns:
Tensor: The solution of a square system of linear equations with a unique solution for input 'x' and 'y'.
Tensor: The solution of a square system of linear equations with a unique solution for input 'x' and 'y'.
Its data type should be the same as that of `x`.
Examples:
.. code-block:: python
# a square system of linear equations:
# 2*X0 + X1 = 9
# X0 + 2*X1 = 8
import paddle
import numpy as np
np_x = np.array([[3, 1],[1, 2]])
np_y = np.array([9, 8])
x = paddle.to_tensor(np_x, dtype="float64")
y = paddle.to_tensor(np_y, dtype="float64")
out = paddle.linalg.solve(x, y)
print(out)
# [2., 3.])
"""
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或