linalg.py 114.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13
#   Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
14

myq406450149's avatar
myq406450149 已提交
15
import numpy as np
Z
Zhang Ting 已提交
16
from ..fluid.layer_helper import LayerHelper
J
Jiabin Yang 已提交
17
from ..framework import _varbase_creator, _dygraph_tracer
H
huangxu96 已提交
18
from ..fluid.data_feeder import check_variable_and_dtype, check_type, check_dtype
Z
zhiboniu 已提交
19
from ..static import Variable
J
Jiabin Yang 已提交
20
from ..fluid.framework import _in_legacy_dygraph, in_dygraph_mode
21
from ..fluid.layers import transpose, cast  # noqa: F401
A
andyjpaddle 已提交
22 23
from ..fluid import layers
import paddle
24 25
from paddle.common_ops_import import core
from paddle.common_ops_import import VarDesc
W
wanghuancoder 已提交
26
from paddle import _C_ops
27

28 29
__all__ = []

30

S
ShenLiang 已提交
31
def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
32
    """
33 34
    Applies matrix multiplication to two tensors. `matmul` follows
    the complete broadcast rules,
S
ShenLiang 已提交
35
    and its behavior is consistent with `np.matmul`.
S
swtkiwi 已提交
36

S
ShenLiang 已提交
37 38
    Currently, the input tensors' number of dimensions can be any, `matmul` can be used to
    achieve the `dot`, `matmul` and `batchmatmul`.
39 40 41 42 43

    The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
    flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:

    - If a transpose flag is specified, the last two dimensions of the tensor
44 45
      are transposed. If the tensor is ndim-1 of shape, the transpose is invalid. If the tensor
      is ndim-1 of shape :math:`[D]`, then for :math:`x` it is treated as :math:`[1, D]`, whereas
S
ShenLiang 已提交
46 47 48 49 50 51 52 53
      for :math:`y` it is the opposite: It is treated as :math:`[D, 1]`.

    The multiplication behavior depends on the dimensions of `x` and `y`. Specifically:

    - If both tensors are 1-dimensional, the dot product result is obtained.

    - If both tensors are 2-dimensional, the matrix-matrix product is obtained.

54 55
    - If the `x` is 1-dimensional and the `y` is 2-dimensional,
      a `1` is prepended to its dimension in order to conduct the matrix multiply.
S
ShenLiang 已提交
56
      After the matrix multiply, the prepended dimension is removed.
57 58

    - If the `x` is 2-dimensional and `y` is 1-dimensional,
S
ShenLiang 已提交
59 60
      the matrix-vector product is obtained.

61 62 63 64 65 66 67 68 69
    - If both arguments are at least 1-dimensional and at least one argument
      is N-dimensional (where N > 2), then a batched matrix multiply is obtained.
      If the first argument is 1-dimensional, a 1 is prepended to its dimension
      in order to conduct the batched matrix multiply and removed after.
      If the second argument is 1-dimensional, a 1 is appended to its
      dimension for the purpose of the batched matrix multiple and removed after.
      The non-matrix (exclude the last two dimensions) dimensions are
      broadcasted according the broadcast rule.
      For example, if input is a (j, 1, n, m) tensor and the other is a (k, m, p) tensor,
S
ShenLiang 已提交
70
      out will be a (j, k, n, p) tensor.
71 72

    Args:
S
ShenLiang 已提交
73 74
        x (Tensor): The input tensor which is a Tensor.
        y (Tensor): The input tensor which is a Tensor.
75 76 77 78 79 80
        transpose_x (bool): Whether to transpose :math:`x` before multiplication.
        transpose_y (bool): Whether to transpose :math:`y` before multiplication.
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.

    Returns:
S
ShenLiang 已提交
81
        Tensor: The output Tensor.
82 83 84

    Examples:

S
ShenLiang 已提交
85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
    .. code-block:: python

        import paddle
        import numpy as np

        # vector * vector
        x_data = np.random.random([10]).astype(np.float32)
        y_data = np.random.random([10]).astype(np.float32)
        x = paddle.to_tensor(x_data)
        y = paddle.to_tensor(y_data)
        z = paddle.matmul(x, y)
        print(z.numpy().shape)
        # [1]

        # matrix * vector
        x_data = np.random.random([10, 5]).astype(np.float32)
        y_data = np.random.random([5]).astype(np.float32)
        x = paddle.to_tensor(x_data)
        y = paddle.to_tensor(y_data)
        z = paddle.matmul(x, y)
        print(z.numpy().shape)
        # [10]

        # batched matrix * broadcasted vector
        x_data = np.random.random([10, 5, 2]).astype(np.float32)
        y_data = np.random.random([2]).astype(np.float32)
        x = paddle.to_tensor(x_data)
        y = paddle.to_tensor(y_data)
        z = paddle.matmul(x, y)
        print(z.numpy().shape)
        # [10, 5]

        # batched matrix * batched matrix
        x_data = np.random.random([10, 5, 2]).astype(np.float32)
        y_data = np.random.random([10, 2, 5]).astype(np.float32)
        x = paddle.to_tensor(x_data)
        y = paddle.to_tensor(y_data)
        z = paddle.matmul(x, y)
        print(z.numpy().shape)
        # [10, 5, 5]

        # batched matrix * broadcasted matrix
        x_data = np.random.random([10, 1, 5, 2]).astype(np.float32)
        y_data = np.random.random([1, 3, 2, 5]).astype(np.float32)
        x = paddle.to_tensor(x_data)
        y = paddle.to_tensor(y_data)
        z = paddle.matmul(x, y)
        print(z.numpy().shape)
        # [10, 3, 5, 5]
134 135

    """
136 137 138 139 140
    if in_dygraph_mode():
        return _C_ops.final_state_matmul(x, y, transpose_x, transpose_y)

    if _in_legacy_dygraph():
        op_type = 'matmul_v2'
W
wanghuancoder 已提交
141
        op = getattr(_C_ops, op_type)
S
ShenLiang 已提交
142 143
        return op(x, y, 'trans_x', transpose_x, 'trans_y', transpose_y)

144
    attrs = {
S
ShenLiang 已提交
145 146
        'trans_x': transpose_x,
        'trans_y': transpose_y,
147 148 149 150 151
    }

    def __check_input(x, y):
        var_names = {'x': x, 'y': y}
        for name, val in var_names.items():
S
ShenLiang 已提交
152
            check_variable_and_dtype(
153 154 155
                val, name,
                ['float16', 'float32', 'float64', 'complex64', 'complex128'],
                'matmul')
156 157 158

    __check_input(x, y)

S
ShenLiang 已提交
159
    helper = LayerHelper('matmul_v2', **locals())
160 161
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
S
ShenLiang 已提交
162
        type='matmul_v2',
163 164 165 166 167
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs=attrs)
    return out
Z
Zhang Ting 已提交
168 169


myq406450149's avatar
myq406450149 已提交
170
def norm(x, p='fro', axis=None, keepdim=False, name=None):
171
    """
S
swtkiwi 已提交
172

173 174 175
    Returns the matrix norm (Frobenius) or vector norm (the 1-norm, the Euclidean
    or 2-norm, and in general the p-norm for p > 0) of a given tensor.

176 177 178 179 180 181
    .. note::
        This norm API is different from `numpy.linalg.norm`.
        This api supports high-order input tensors (rank >= 3), and certain axis need to be pointed out to calculate the norm.
        But `numpy.linalg.norm` only supports 1-D vector or 2-D matrix as input tensor.
        For p-order matrix norm, this api actually treats matrix as a flattened vector to calculate the vector norm, NOT REAL MATRIX NORM.

182
    Args:
myq406450149's avatar
myq406450149 已提交
183
        x (Tensor): The input tensor could be N-D tensor, and the input data
184
            type could be float32 or float64.
myq406450149's avatar
myq406450149 已提交
185
        p (float|string, optional): Order of the norm. Supported values are `fro`, `0`, `1`, `2`,
186
            `inf`, `-inf` and any positive real number yielding the corresponding p-norm. Not supported: ord < 0 and nuclear norm.
myq406450149's avatar
myq406450149 已提交
187
            Default value is `fro`.
myq406450149's avatar
myq406450149 已提交
188 189
        axis (int|list|tuple, optional): The axis on which to apply norm operation. If axis is int
            or list(int)/tuple(int)  with only one element, the vector norm is computed over the axis.
190
            If `axis < 0`, the dimension to norm operation is rank(input) + axis.
myq406450149's avatar
myq406450149 已提交
191
            If axis is a list(int)/tuple(int) with two elements, the matrix norm is computed over the axis.
myq406450149's avatar
myq406450149 已提交
192
            Defalut value is `None`.
193 194 195 196 197 198 199 200
        keepdim (bool, optional): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have fewer dimension
            than the :attr:`input` unless :attr:`keepdim` is true, default
            value is False.
        name (str, optional): The default value is None. Normally there is no need for
            user to set this property. For more information, please refer to :ref:`api_guide_Name`.

    Returns:
myq406450149's avatar
myq406450149 已提交
201
        Tensor: results of norm operation on the specified axis of input tensor,
202
        it's data type is the same as input's Tensor.
203

204 205
    Examples:
        .. code-block:: python
206

207
            import paddle
myq406450149's avatar
myq406450149 已提交
208 209 210 211 212 213 214 215
            import numpy as np
            shape=[2, 3, 4]
            np_input = np.arange(24).astype('float32') - 12
            np_input = np_input.reshape(shape)
            x = paddle.to_tensor(np_input)
            #[[[-12. -11. -10.  -9.] [ -8.  -7.  -6.  -5.] [ -4.  -3.  -2.  -1.]]
            # [[  0.   1.   2.   3.] [  4.   5.   6.   7.] [  8.   9.  10.  11.]]]

216
            # compute frobenius norm along last two dimensions.
217
            out_fro = paddle.linalg.norm(x, p='fro', axis=[0,1])
myq406450149's avatar
myq406450149 已提交
218 219
            # out_fro.numpy() [17.435596 16.911535 16.7332   16.911535]

220
            # compute 2-order vector norm along last dimension.
221
            out_pnorm = paddle.linalg.norm(x, p=2, axis=-1)
myq406450149's avatar
myq406450149 已提交
222 223 224 225
            #out_pnorm.numpy(): [[21.118711  13.190906   5.477226]
            #                    [ 3.7416575 11.224972  19.131126]]

            # compute 2-order  norm along [0,1] dimension.
226
            out_pnorm = paddle.linalg.norm(x, p=2, axis=[0,1])
myq406450149's avatar
myq406450149 已提交
227 228 229
            #out_pnorm.numpy(): [17.435596 16.911535 16.7332   16.911535]

            # compute inf-order  norm
230
            out_pnorm = paddle.linalg.norm(x, p=np.inf)
myq406450149's avatar
myq406450149 已提交
231
            #out_pnorm.numpy()  = [12.]
232
            out_pnorm = paddle.linalg.norm(x, p=np.inf, axis=0)
myq406450149's avatar
myq406450149 已提交
233 234 235
            #out_pnorm.numpy(): [[12. 11. 10. 9.] [8. 7. 6. 7.] [8. 9. 10. 11.]]

            # compute -inf-order  norm
236
            out_pnorm = paddle.linalg.norm(x, p=-np.inf)
myq406450149's avatar
myq406450149 已提交
237
            #out_pnorm.numpy(): [0.]
238
            out_pnorm = paddle.linalg.norm(x, p=-np.inf, axis=0)
myq406450149's avatar
myq406450149 已提交
239
            #out_pnorm.numpy(): [[0. 1. 2. 3.] [4. 5. 6. 5.] [4. 3. 2. 1.]]
240 241
    """

myq406450149's avatar
myq406450149 已提交
242
    def frobenius_norm(input, dim=None, keepdim=False, name=None):
243 244 245 246 247 248 249 250 251 252 253
        """
        The frobenius norm OP is to calculate the frobenius norm of certain two dimensions of Tensor `input`.
        Args:
          input (Variable): Tensor, data type float32, float64.
          dim (list, optional): None for last two dimensions.
          keepdim (bool, optional): Whether keep the dimensions as the `input`, Default False.
        """
        if dim is not None and not (isinstance(dim, list) and len(dim) == 2):
            raise ValueError(
                "The dim of frobenius norm op should be None or two elements list!"
            )
Z
zhiboniu 已提交
254
        if paddle.in_dynamic_mode():
myq406450149's avatar
myq406450149 已提交
255
            if dim is None:
W
wanghuancoder 已提交
256 257 258 259
                return _C_ops.frobenius_norm(input, 'keep_dim', keepdim,
                                             'reduce_all', True)
            return _C_ops.frobenius_norm(input, 'dim', dim, 'keep_dim', keepdim,
                                         'reduce_all', False)
myq406450149's avatar
myq406450149 已提交
260 261
        attrs = {'dim': dim, 'keep_dim': keepdim, 'reduce_all': False}
        if dim is None:
262 263 264 265 266
            attrs['reduce_all'] = True
        check_variable_and_dtype(input, 'input', ['float32', 'float64'],
                                 'frobenius_norm')

        helper = LayerHelper('frobenius_norm', **locals())
myq406450149's avatar
myq406450149 已提交
267 268
        out = helper.create_variable_for_type_inference(
            dtype=helper.input_dtype())
269 270 271 272 273 274 275 276 277 278 279 280

        helper.append_op(
            type='frobenius_norm',
            inputs={'X': input},
            outputs={'Out': out},
            attrs=attrs)
        return out

    def vector_norm(input,
                    porder=None,
                    axis=None,
                    keepdim=False,
myq406450149's avatar
myq406450149 已提交
281
                    asvector=False,
282 283 284 285 286 287 288 289 290
                    name=None):
        """
        Calculate the p-order vector norm for certain  dimension of Tensor `input`.
        Args:
          input (Variable): Tensor, data type float32, float64.
          porder (float, optional): None for porder=2.0.
          axis (int, optional): None for last dimension.
          keepdim (bool, optional): Whether keep the dimensions as the `input`, Default False.
        """
Z
zhiboniu 已提交
291
        if paddle.in_dynamic_mode():
myq406450149's avatar
myq406450149 已提交
292
            if axis is None: axis = -1
W
wanghuancoder 已提交
293 294
            return _C_ops.p_norm(input, 'porder', porder, 'axis', axis,
                                 'keepdim', keepdim, 'asvector', asvector)
295 296 297 298
        if porder is not None:
            check_type(porder, 'porder', (float, int), 'p_norm')
        if axis is not None:
            check_type(axis, 'axis', (int), 'p_norm')
myq406450149's avatar
myq406450149 已提交
299 300 301
        check_variable_and_dtype(input, 'input', ['float32', 'float64'],
                                 'p_norm')

302 303 304 305
        attrs = {
            'axis': axis if axis is not None else -1,
            'porder': float(porder) if porder is not None else 2.0,
            'keepdim': keepdim,
myq406450149's avatar
myq406450149 已提交
306
            'asvector': asvector,
307 308 309
            'epsilon': 1e-12,
        }
        helper = LayerHelper('p_norm', **locals())
myq406450149's avatar
myq406450149 已提交
310 311
        out = helper.create_variable_for_type_inference(
            dtype=helper.input_dtype())
312 313 314 315 316 317 318 319

        helper.append_op(
            type='p_norm',
            inputs={'X': input},
            outputs={'Out': out},
            attrs=attrs)
        return out

myq406450149's avatar
myq406450149 已提交
320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348
    def inf_norm(input,
                 porder=None,
                 axis=axis,
                 keepdim=False,
                 asvector=False,
                 name=None):
        helper = LayerHelper('frobenius_norm', **locals())
        out = helper.create_variable_for_type_inference(
            dtype=helper.input_dtype())
        helper.append_op(type='abs', inputs={'X': input}, outputs={'Out': out})
        reduce_out = helper.create_variable_for_type_inference(
            dtype=helper.input_dtype())

        reduce_all = True if axis == None or axis == [] or asvector == True else False
        axis = axis if axis != None and axis != [] else [0]

        reduce_type = 'reduce_max' if porder == np.float(
            'inf') else 'reduce_min'
        helper.append_op(
            type=reduce_type,
            inputs={'X': out},
            outputs={'Out': reduce_out},
            attrs={'dim': axis,
                   'keep_dim': keepdim,
                   'reduce_all': reduce_all})

        return reduce_out

    def p_matrix_norm(input, porder=1., axis=axis, keepdim=False, name=None):
349 350 351 352
        """
        NOTE:
            This function actually treats the matrix as flattened vector to calculate vector norm instead of matrix norm.
        """
myq406450149's avatar
myq406450149 已提交
353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386
        block = LayerHelper('norm', **locals())
        out = block.create_variable_for_type_inference(
            dtype=block.input_dtype())
        abs_out = block.create_variable_for_type_inference(
            dtype=block.input_dtype())
        block.append_op(
            type='abs', inputs={'X': input}, outputs={'Out': abs_out})
        pow_out = block.create_variable_for_type_inference(
            dtype=block.input_dtype())

        block.append_op(
            type='pow',
            inputs={'X': abs_out},
            outputs={'Out': pow_out},
            attrs={'factor': porder})
        sum_out = block.create_variable_for_type_inference(
            dtype=block.input_dtype())
        block.append_op(
            type='reduce_sum',
            inputs={'X': pow_out},
            outputs={'Out': sum_out},
            attrs={
                'dim': axis,
                'keep_dim': keepdim,
                'reduce_all': True if axis is None else False
            })
        porder
        block.append_op(
            type='pow',
            inputs={'X': sum_out},
            outputs={'Out': out},
            attrs={'factor': float(1. / porder)})
        return out

387 388 389
    if axis is None and p is not None:
        if isinstance(p, str):
            if p == "fro":
myq406450149's avatar
myq406450149 已提交
390
                return frobenius_norm(x, dim=axis, keepdim=keepdim, name=name)
391 392 393 394 395
            else:
                raise ValueError(
                    "only valid string values are 'fro', found {}".format(p))
        elif isinstance(p, (int, float)):
            return vector_norm(
myq406450149's avatar
myq406450149 已提交
396 397 398 399 400 401
                x,
                porder=p,
                axis=axis,
                keepdim=keepdim,
                asvector=True,
                name=name)
402 403 404 405
        else:
            raise ValueError("only valid p type is string or float, found {}".
                             format(type(p)))

myq406450149's avatar
myq406450149 已提交
406 407
    if isinstance(axis, tuple):
        axis = list(axis)
408 409 410 411 412
    if isinstance(axis, list) and len(axis) == 1:
        axis = axis[0]

    #calculate vector norm, where axis is int or list with only one integer
    if isinstance(axis, int):
myq406450149's avatar
myq406450149 已提交
413 414 415 416 417 418 419 420 421 422 423 424 425 426
        if isinstance(p, str):
            if p == "fro":
                return vector_norm(
                    x,
                    porder=2,
                    axis=axis,
                    keepdim=keepdim,
                    asvector=False,
                    name=name)

            else:
                raise ValueError(
                    "only valid string values are 'fro', found {}".format(p))
        elif isinstance(p, (int, float)):
427
            return vector_norm(
myq406450149's avatar
myq406450149 已提交
428 429 430 431 432 433
                x,
                axis=axis,
                porder=p,
                keepdim=keepdim,
                asvector=False,
                name=name)
434 435 436 437 438 439 440
        else:
            raise ValueError(
                "unspport p for p-order vector norm. except float, found {}".
                format(p))
    #calculate matrix norm, where axis is list with two integers
    elif isinstance(axis, list) and len(axis) == 2:
        if p == "fro":
myq406450149's avatar
myq406450149 已提交
441 442 443
            return frobenius_norm(x, dim=axis, keepdim=keepdim, name=name)
        elif p == np.inf or p == -np.inf:
            return inf_norm(x, porder=p, axis=axis, keepdim=keepdim, name=name)
myq406450149's avatar
myq406450149 已提交
444 445 446 447
        elif p == 0:
            raise ValueError(
                "just suport axis type int or list (length of list <=1) if p = 0, found {}".
                format(axis))
448
        else:
myq406450149's avatar
myq406450149 已提交
449 450
            return p_matrix_norm(
                x, porder=p, axis=axis, keepdim=keepdim, name=name)
451 452 453 454 455 456
    else:
        raise ValueError(
            "except axis type int or list (length of list <=2), found {}".
            format(axis))


457
def dist(x, y, p=2, name=None):
458
    r"""
S
swtkiwi 已提交
459

Z
Zhang Ting 已提交
460
    This OP returns the p-norm of (x - y). It is not a norm in a strict sense, only as a measure
461 462
    of distance. The shapes of x and y must be broadcastable. The definition is as follows, for
    details, please refer to the `numpy's broadcasting <https://docs.scipy.org/doc/numpy/user/basics.broadcasting.html>`_:
Z
Zhang Ting 已提交
463

464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486
    - Each input has at least one dimension.
    - Match the two input dimensions from back to front, the dimension sizes must either be equal, one of them is 1, or one of them does not exist.

    Where, z = x - y, the shapes of x and y are broadcastable, then the shape of z can be
    obtained as follows:

    1. If the number of dimensions of x and y are not equal, prepend 1 to the dimensions of the
    tensor with fewer dimensions.

    For example, The shape of x is [8, 1, 6, 1], the shape of y is [7, 1, 5], prepend 1 to the
    dimension of y.

    x (4-D Tensor):  8 x 1 x 6 x 1

    y (4-D Tensor):  1 x 7 x 1 x 5

    2. Determine the size of each dimension of the output z: choose the maximum value from the
    two input dimensions.

    z (4-D Tensor):  8 x 7 x 6 x 5

    If the number of dimensions of the two inputs are the same, the size of the output can be
    directly determined in step 2. When p takes different values, the norm formula is as follows:
Z
Zhang Ting 已提交
487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512

    When p = 0, defining $0^0=0$, the zero-norm of z is simply the number of non-zero elements of z.

    .. math::

        ||z||_{0}=\lim_{p \\rightarrow 0}\sum_{i=1}^{m}|z_i|^{p}

    When p = inf, the inf-norm of z is the maximum element of z.

    .. math::

        ||z||_\infty=\max_i |z_i|

    When p = -inf, the negative-inf-norm of z is the minimum element of z.

    .. math::

        ||z||_{-\infty}=\min_i |z_i|

    Otherwise, the p-norm of z follows the formula,

    .. math::

        ||z||_{p}=(\sum_{i=1}^{m}|z_i|^p)^{\\frac{1}{p}}

    Args:
513 514
        x (Tensor): 1-D to 6-D Tensor, its data type is float32 or float64.
        y (Tensor): 1-D to 6-D Tensor, its data type is float32 or float64.
Z
Zhang Ting 已提交
515 516 517
        p (float, optional): The norm to be computed, its data type is float32 or float64. Default: 2.

    Returns:
518
        Tensor: Tensor that is the p-norm of (x - y).
Z
Zhang Ting 已提交
519 520 521 522 523 524 525

    Examples:
        .. code-block:: python

            import paddle
            import numpy as np

526 527 528 529
            x = paddle.to_tensor(np.array([[3, 3],[3, 3]]), "float32")
            y = paddle.to_tensor(np.array([[3, 3],[3, 1]]), "float32")
            out = paddle.dist(x, y, 0)
            print(out) # out = [1.]
Z
Zhang Ting 已提交
530

531 532
            out = paddle.dist(x, y, 2)
            print(out) # out = [2.]
Z
Zhang Ting 已提交
533

534 535
            out = paddle.dist(x, y, float("inf"))
            print(out) # out = [2.]
Z
Zhang Ting 已提交
536

537 538
            out = paddle.dist(x, y, float("-inf"))
            print(out) # out = [0.]
Z
Zhang Ting 已提交
539 540 541 542 543 544 545 546 547 548 549 550 551
    """
    check_variable_and_dtype(x, 'dtype', ['float32', 'float64'], 'dist')
    check_variable_and_dtype(y, 'dtype', ['float32', 'float64'], 'dist')
    check_type(p, 'p', (float, int), 'dist')
    helper = LayerHelper("dist", **locals())
    out = helper.create_variable_for_type_inference(x.dtype)

    inputs = {"X": [x], "Y": [y]}
    outputs = {'Out': [out]}
    attrs = {"p": float(p)}
    helper.append_op(
        type='dist', inputs=inputs, outputs={'Out': out}, attrs=attrs)
    return out
L
liuwei1031 已提交
552 553


554 555 556 557 558 559
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:
560 561
        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``.
562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615
            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`.
        name (str, optional): The default value is `None`. Normally there is no need for
            user to set this property. For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor: computing results of condition number, its data type is the same as input Tensor ``x``.

    Examples:
        .. code-block:: python

            import paddle
            import numpy as np

            x = paddle.to_tensor([[1., 0, -1], [0, 1, 0], [1, 0, 1]])

            # compute conditional number when p is None
            out = paddle.linalg.cond(x)
            # out.numpy() [1.4142135]

            # compute conditional number when order of the norm is 'fro'
            out_fro = paddle.linalg.cond(x, p='fro')
            # out_fro.numpy() [3.1622777]

            # compute conditional number when order of the norm is 'nuc'
            out_nuc = paddle.linalg.cond(x, p='nuc')
            # out_nuc.numpy() [9.2426405]

            # compute conditional number when order of the norm is 1
            out_1 = paddle.linalg.cond(x, p=1)
            # out_1.numpy() [2.]

            # compute conditional number when order of the norm is -1
            out_minus_1 = paddle.linalg.cond(x, p=-1)
            # out_minus_1.numpy() [1.]

            # compute conditional number when order of the norm is 2
            out_2 = paddle.linalg.cond(x, p=2)
            # out_2.numpy() [1.4142135]

            # compute conditional number when order of the norm is -1
            out_minus_2 = paddle.linalg.cond(x, p=-2)
            # out_minus_2.numpy() [0.70710677]

            # compute conditional number when order of the norm is inf
            out_inf = paddle.linalg.cond(x, p=np.inf)
            # out_inf.numpy() [2.]

            # compute conditional number when order of the norm is -inf
            out_minus_inf = paddle.linalg.cond(x, p=-np.inf)
            # out_minus_inf.numpy() [1.]

            a = paddle.to_tensor(np.random.randn(2, 4, 4).astype('float32'))
616
            # a.numpy()
617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650
            # [[[ 0.14063153 -0.996288    0.7996131  -0.02571543]
            #   [-0.16303636  1.5534962  -0.49919784 -0.04402903]
            #   [-1.1341571  -0.6022629   0.5445269   0.29154757]
            #   [-0.16816919 -0.30972657  1.7521842  -0.5402487 ]]
            #  [[-0.58081484  0.12402827  0.7229862  -0.55046535]
            #   [-0.15178485 -1.1604939   0.75810957  0.30971205]
            #   [-0.9669573   1.0940945  -0.27363303 -0.35416734]
            #   [-1.216529    2.0018666  -0.7773689  -0.17556527]]]
            a_cond_fro = paddle.linalg.cond(a, p='fro')
            # a_cond_fro.numpy()  [31.572273 28.120834]

            b = paddle.to_tensor(np.random.randn(2, 3, 4).astype('float64'))
            # b.numpy()
            # [[[ 1.61707487  0.46829144  0.38130416  0.82546736]
            #   [-1.72710298  0.08866375 -0.62518804  0.16128892]
            #   [-0.02822879 -1.67764516  0.11141444  0.3220113 ]]
            #  [[ 0.22524372  0.62474921 -0.85503233 -1.03960523]
            #   [-0.76620689  0.56673047  0.85064753 -0.45158196]
            #   [ 1.47595418  2.23646462  1.5701758   0.10497519]]]
            b_cond_2 = paddle.linalg.cond(b, p=2)
            # b_cond_2.numpy()  [3.30064451 2.51976252]

    """

    def mat_norm(input, porder=1., axis=None):
        """
        NOTE:
            Calculate the matrix norm of a square matrix or batches of square matrices,
            when porder is in (1, -1, inf, -inf)
        """
        reduce_all = True if axis is None or axis == [] else False
        axis = axis if axis != None and axis != [] else [0]
        keepdim = False

Z
zhiboniu 已提交
651
        if paddle.in_dynamic_mode():
652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707
            abs_out = _C_ops.abs(input)
            sum_out = _C_ops.reduce_sum(abs_out, 'dim', axis, 'keepdim',
                                        keepdim, 'reduce_all', reduce_all)
            if porder == 1 or porder == np.inf:
                return _C_ops.reduce_max(sum_out, 'dim', [-1], 'keepdim',
                                         keepdim, 'reduce_all', reduce_all)
            if porder == -1 or porder == -np.inf:
                return _C_ops.reduce_min(sum_out, 'dim', [-1], 'keepdim',
                                         keepdim, 'reduce_all', reduce_all)

        block = LayerHelper('norm', **locals())
        abs_out = block.create_variable_for_type_inference(
            dtype=block.input_dtype())
        sum_out = block.create_variable_for_type_inference(
            dtype=block.input_dtype())
        out = block.create_variable_for_type_inference(
            dtype=block.input_dtype())
        block.append_op(
            type='abs', inputs={'X': input}, outputs={'Out': abs_out})
        block.append_op(
            type='reduce_sum',
            inputs={'X': abs_out},
            outputs={'Out': sum_out},
            attrs={'dim': axis,
                   'keep_dim': keepdim,
                   'reduce_all': reduce_all})
        if porder == 1 or porder == np.inf:
            block.append_op(
                type='reduce_max',
                inputs={'X': sum_out},
                outputs={'Out': out},
                attrs={
                    'dim': [-1],
                    'keep_dim': keepdim,
                    'reduce_all': reduce_all
                })
        if porder == -1 or porder == -np.inf:
            block.append_op(
                type='reduce_min',
                inputs={'X': sum_out},
                outputs={'Out': out},
                attrs={
                    'dim': [-1],
                    'keep_dim': keepdim,
                    'reduce_all': reduce_all
                })
        return out

    def fro_norm(input, porder=2, axis=[-1]):
        """
        NOTE:
            Calculate the frobenius norm of a square matrix or batches of square matrices.
        """
        reduce_all = True if axis is None or axis == [] else False
        keepdim = False

Z
zhiboniu 已提交
708
        if paddle.in_dynamic_mode():
709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761
            pow_out = _C_ops.pow(input, 'factor', porder)
            sum_out_1 = _C_ops.reduce_sum(pow_out, 'dim', axis, 'keepdim',
                                          keepdim, 'reduce_all', reduce_all)
            sum_out_2 = _C_ops.reduce_sum(sum_out_1, 'dim', axis, 'keepdim',
                                          keepdim, 'reduce_all', reduce_all)
            return _C_ops.pow(sum_out_2, 'factor', float(1. / porder))

        block = LayerHelper('norm', **locals())
        pow_out = block.create_variable_for_type_inference(
            dtype=block.input_dtype())
        sum_out_1 = block.create_variable_for_type_inference(
            dtype=block.input_dtype())
        sum_out_2 = block.create_variable_for_type_inference(
            dtype=block.input_dtype())
        out = block.create_variable_for_type_inference(
            dtype=block.input_dtype())
        block.append_op(
            type='pow',
            inputs={'X': input},
            outputs={'Out': pow_out},
            attrs={'factor': porder})
        block.append_op(
            type='reduce_sum',
            inputs={'X': pow_out},
            outputs={'Out': sum_out_1},
            attrs={'dim': axis,
                   'keep_dim': keepdim,
                   'reduce_all': reduce_all})
        block.append_op(
            type='reduce_sum',
            inputs={'X': sum_out_1},
            outputs={'Out': sum_out_2},
            attrs={'dim': axis,
                   'keep_dim': keepdim,
                   'reduce_all': reduce_all})
        block.append_op(
            type='pow',
            inputs={'X': sum_out_2},
            outputs={'Out': out},
            attrs={'factor': float(1. / porder)})
        return out

    def svd_norm(input, porder, axis=[-1]):
        """
        NOTE:
            Calculate the matrix norm, which is related to singular values, of a matrix
            or batches of matrices, including nuclear norm, 2-norm and (-2)-norm.
        """
        reduce_all = True if axis is None or axis == [] else False
        keepdim = False

        u, s, vh = svd(input, full_matrices=False)

Z
zhiboniu 已提交
762
        if paddle.in_dynamic_mode():
763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828
            if porder == "nuc":
                return _C_ops.reduce_sum(s, 'dim', axis, 'keepdim', keepdim,
                                         'reduce_all', reduce_all)
            max_out = _C_ops.reduce_max(s, 'dim', axis, 'keepdim', keepdim,
                                        'reduce_all', reduce_all)
            min_out = _C_ops.reduce_min(s, 'dim', axis, 'keepdim', keepdim,
                                        'reduce_all', reduce_all)
            if porder == 2:
                return _C_ops.elementwise_div(max_out, min_out, 'aixs', axis,
                                              'use_mkldnn', False)
            if porder == -2:
                return _C_ops.elementwise_div(min_out, max_out, 'aixs', axis,
                                              'use_mkldnn', False)

        block = LayerHelper('norm', **locals())
        out = block.create_variable_for_type_inference(
            dtype=block.input_dtype())
        if porder == "nuc":
            block.append_op(
                type='reduce_sum',
                inputs={'X': s},
                outputs={'Out': out},
                attrs={
                    'dim': axis,
                    'keep_dim': keepdim,
                    'reduce_all': reduce_all
                })
            return out
        max_out = block.create_variable_for_type_inference(
            dtype=block.input_dtype())
        min_out = block.create_variable_for_type_inference(
            dtype=block.input_dtype())
        block.append_op(
            type='reduce_max',
            inputs={'X': s},
            outputs={'Out': max_out},
            attrs={'dim': axis,
                   'keep_dim': keepdim,
                   'reduce_all': reduce_all})
        block.append_op(
            type='reduce_min',
            inputs={'X': s},
            outputs={'Out': min_out},
            attrs={'dim': axis,
                   'keep_dim': keepdim,
                   'reduce_all': reduce_all})
        if porder == 2:
            block.append_op(
                type='elementwise_div',
                inputs={'X': max_out,
                        'Y': min_out},
                outputs={'Out': out},
                attrs={'aixs': axis,
                       'use_mkldnn': False})
            return out
        if porder == -2:
            block.append_op(
                type='elementwise_div',
                inputs={'X': min_out,
                        'Y': max_out},
                outputs={'Out': out},
                attrs={'aixs': axis,
                       'use_mkldnn': False})
            return out

    def empty_tensor(input, shape):
Z
zhiboniu 已提交
829
        if paddle.in_dynamic_mode():
830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870
            return input.reshape(shape)
        raise ValueError("only support x is nonempty tensor in static mode")

    x_shape = list(x.shape)
    if not len(x_shape) >= 2:
        raise ValueError("input should be a matrix or batches of matrices, " +
                         "but the dimention of received input is {}".format(
                             len(x_shape)))
    if p == None:
        p = 2
    x_size = 0 if (0 in x_shape) else 1
    if p in ("fro", "nuc", 1, -1, np.inf, -np.inf):
        if x_shape[len(x_shape) - 1] == x_shape[len(x_shape) - 2]:
            if x_size == 0:
                return empty_tensor(x, x_shape[:-2])
            x_inv = x.inverse()
            if p == "fro":
                return fro_norm(x) * fro_norm(x_inv)
            if p == "nuc":
                return svd_norm(x, p) * svd_norm(x_inv, p)
            if p in (1, -1):
                return mat_norm(
                    x, porder=p, axis=[-2]) * mat_norm(
                        x_inv, porder=p, axis=[-2])
            if p in (np.inf, -np.inf):
                return mat_norm(
                    x, porder=p, axis=[-1]) * mat_norm(
                        x_inv, porder=p, axis=[-1])
        else:
            raise ValueError("only support p is {} when input is a ".format(p) +
                             "square matrix or batches of square matrices")
    elif p in (2, -2):
        if x_size == 0:
            return empty_tensor(x, x_shape[:-2])
        return svd_norm(x, porder=p)
    else:
        raise ValueError(
            "unsupported {} for p, only supporting ('fro', 'nuc', ".format(
                p) + "1, -1, 2, -2, inf, -inf) or none")


L
liuwei1031 已提交
871 872 873
def dot(x, y, name=None):
    """
    This operator calculates inner product for vectors.
874

L
liuwei1031 已提交
875
    .. note::
876 877
       Support 1-d and 2-d Tensor. When it is 2d, the first dimension of this matrix
       is the batch dimension, which means that the vectors of multiple batches are dotted.
L
liuwei1031 已提交
878 879

    Parameters:
S
ShenLiang 已提交
880 881
        x(Tensor): 1-D or 2-D ``Tensor``. Its dtype should be ``float32``, ``float64``, ``int32``, ``int64``
        y(Tensor): 1-D or 2-D ``Tensor``. Its dtype soulde be ``float32``, ``float64``, ``int32``, ``int64``
L
liuwei1031 已提交
882 883
        name(str, optional): Name of the output. Default is None. It's used to print debug info for developers. Details: :ref:`api_guide_Name`

884
    Returns:
885
        Tensor: the calculated result Tensor.
886

L
liuwei1031 已提交
887 888 889 890 891 892
    Examples:

    .. code-block:: python

        import paddle
        import numpy as np
893 894 895

        x_data = np.random.uniform(0.1, 1, [10]).astype(np.float32)
        y_data = np.random.uniform(1, 3, [10]).astype(np.float32)
S
ShenLiang 已提交
896 897
        x = paddle.to_tensor(x_data)
        y = paddle.to_tensor(y_data)
898
        z = paddle.dot(x, y)
899
        print(z)
L
liuwei1031 已提交
900 901 902

    """
    op_type = 'dot'
903
    # skip var type check in dygraph mode to improve efficiency
Z
zhiboniu 已提交
904
    if paddle.in_dynamic_mode():
W
wanghuancoder 已提交
905
        op = getattr(_C_ops, op_type)
906 907
        return op(x, y)

L
liuwei1031 已提交
908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925
    assert x is not None, 'x cannot be None in {}'.format(op_type)
    assert y is not None, 'y cannot be None in {}'.format(op_type)

    check_variable_and_dtype(x, 'x', ['float32', 'float64', 'int32', 'int64'],
                             op_type)
    check_variable_and_dtype(y, 'y', ['float32', 'float64', 'int32', 'int64'],
                             op_type)

    helper = LayerHelper(op_type, **locals())
    if name is None:
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)
    helper.append_op(
        type="dot", inputs={'X': x,
                            'Y': y}, attrs={}, outputs={"Out": out})
    return out
926 927


Z
zhiboniu 已提交
928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040
def cov(x, rowvar=True, ddof=True, fweights=None, aweights=None, name=None):
    """
    Estimate the covariance matrix of the input variables, given data and weights.

    A covariance matrix is a square matrix, indicate the covariance of each pair variables in the input matrix.
    For example, for an N-dimensional samples X=[x1,x2,…xN]T, then the covariance matrix 
    element Cij is the covariance of xi and xj. The element Cii is the variance of xi itself.

    Parameters:
        x(Tensor): A N-D(N<=2) Tensor containing multiple variables and observations. By default, each row of x represents a variable. Also see rowvar below.
        rowvar(Bool, optional): If rowvar is True (default), then each row represents a variable, with observations in the columns. Default: True
        ddof(Bool, optional): If ddof=True will return the unbiased estimate, and ddof=False will return the simple average. Default: True
        fweights(Tensor, optional): 1-D Tensor of integer frequency weights; The number of times each observation vector should be repeated. Default: None
        aweights(Tensor, optional): 1-D Tensor of observation vector weights. How important of the observation vector, larger data means this element is more important. Default: None
        name(str, optional): Name of the output. Default is None. It's used to print debug info for developers. Details: :ref:`api_guide_Name`

    Returns:
        Tensor: The covariance matrix Tensor of the variables.

    Examples:

    .. code-block:: python

        import paddle

        xt = paddle.rand((3,4))
        paddle.linalg.cov(xt)

        '''
        Tensor(shape=[3, 3], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
            [[0.07918842, 0.06127326, 0.01493049],
                [0.06127326, 0.06166256, 0.00302668],
                [0.01493049, 0.00302668, 0.01632146]])
        '''
    """
    op_type = 'cov'
    if len(x.shape) > 2 or len(x.shape) < 1:
        raise ValueError(
            "Input(x) only support N-D (1<=N<=2) tensor in cov, but received "
            "length of Input(input) is %s." % len(x.shape))
    check_variable_and_dtype(x, 'dtype', ['float32', 'float64'], 'cov')
    nx = x
    if len(x.shape) == 1:
        nx = x.reshape((1, -1))
    if not rowvar and nx.shape[0] != 1:
        nx = nx.t()
    w = None
    observation_num = nx.shape[1]
    if fweights is not None:
        w = fweights.astype(nx.dtype)
        if len(w.shape) > 1:
            raise ValueError(
                "Input(fweights) only support N-D (N<=1) tensor in cov, but received "
                "shape of Input(input) is %s." % len(fweights.shape))
        if fweights.shape[0] != observation_num:
            raise ValueError(
                "The number of Input(fweights) should equal to x's dim[1]: {}, but received "
                "size of Input(fweights) is {}.".format(observation_num,
                                                        fweights.shape[0]))
        if fweights.min() < 0:
            raise ValueError(
                "The value of Input(fweights) cannot be negtive, but received "
                "min of Input(fweights) is {}.".format(fweights.min()))
        if not paddle.all(fweights == paddle.round(fweights.astype('float64'))):
            raise ValueError("Input(fweights) must be integer ")

    if aweights is not None:
        aw = aweights.astype(nx.dtype)
        if len(aw.shape) > 1:
            raise ValueError(
                "Input(aweights) only support N-D (N<=1) tensor in cov, but received "
                "length of Input(input) is %s." % len(aweights.shape))
        check_variable_and_dtype(aweights, 'dtype', ['float32', 'float64'],
                                 'cov')
        if aweights.shape[0] != observation_num:
            raise ValueError(
                "The number of Input(aweights) should equal to x's dim[1]: {}, but received "
                "size of Input(aweights) is {}.".format(observation_num,
                                                        aweights.shape[0]))
        if aweights.min() < 0:
            raise ValueError(
                "The value of Input(aweights) cannot be negtive, but received "
                "min of Input(aweights) is {}.".format(aweights.min()))
        if w is not None:
            w = w * aw
        else:
            w = aw

    w_sum = paddle.to_tensor(observation_num, dtype=nx.dtype)
    if fweights is not None or aweights is not None:
        w_sum = w.sum()
        if w_sum.item() == 0:
            raise ValueError("The sum of weights is zero, can't be normalized.")

    if w is not None:
        nx_w = nx * w
        avg = (nx_w).sum(axis=1) / w_sum
    else:
        avg = nx.sum(axis=1) / w_sum
        nx_w = nx

    if w is not None and aweights is not None and ddof == True:
        norm_factor = w_sum - (w * aweights).sum() / w_sum
    else:
        norm_factor = w_sum - ddof
    if norm_factor <= 0:
        norm_factor = paddle.to_tensor(0, dtype=nx.dtype)
    nx = nx - avg.unsqueeze(1)
    xxt = paddle.mm(nx, nx_w.t().conj())
    cov = paddle.divide(xxt, norm_factor).squeeze()
    return cov


1041 1042
def t(input, name=None):
    """
1043 1044
    Transpose <=2-D tensor.
    0-D and 1-D tensors are returned as it is and 2-D tensor is equal to
1045
    the paddle.transpose function which perm dimensions set 0 and 1.
1046

1047
    Args:
1048
        input (Tensor): The input Tensor. It is a N-D (N<=2) Tensor of data types float16, float32, float64, int32.
1049
        name(str, optional): The default value is None.  Normally there is no need for
1050 1051
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`
    Returns:
1052
        Tensor: A transposed n-D Tensor, with data type being float16, float32, float64, int32, int64.
1053

1054
    For Example:
1055

1056
        .. code-block:: text
1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072

             # Example 1 (0-D tensor)
             x = tensor([0.79])
             paddle.t(x) = tensor([0.79])

             # Example 2 (1-D tensor)
             x = tensor([0.79, 0.84, 0.32])
             paddle.t(x) = tensor([0.79, 0.84, 0.32])

             # Example 3 (2-D tensor)
             x = tensor([0.79, 0.84, 0.32],
                        [0.64, 0.14, 0.57])
             paddle.t(x) = tensor([0.79, 0.64],
                                  [0.84, 0.14],
                                  [0.32, 0.57])

1073
     Examples:
1074

1075
        .. code-block:: python
1076

1077
            import paddle
1078
            x = paddle.ones(shape=[2, 3], dtype='int32')
1079
            x_transposed = paddle.t(x)
1080 1081
            print(x_transposed.shape)
            # [3, 2]
1082 1083 1084 1085 1086 1087
    """
    if len(input.shape) > 2:
        raise ValueError(
            "Input(input) only support N-D (N<=2) tensor, but received "
            "length of Input(input) is %s. Perhaps you can use paddle."
            "tensor.transpose() instead." % len(input.shape))
Z
zhiboniu 已提交
1088
    if paddle.in_dynamic_mode():
1089 1090 1091 1092
        if len(input.shape) == 1:
            return input
        # 2-D tensor
        perm = [1, 0]
W
wanghuancoder 已提交
1093
        out, _ = _C_ops.transpose2(input, 'axis', perm)
1094 1095 1096
        return out

    check_variable_and_dtype(
1097 1098
        input, 'input', ['float16', 'float32', 'float64', 'int32',
                         'int64'], 'transpose')
1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112

    helper = LayerHelper('t', **locals())
    out = helper.create_variable_for_type_inference(input.dtype)
    input_shape = helper.create_variable_for_type_inference(input.dtype)
    if len(input.shape) == 1:
        out = input
    else:
        helper.append_op(
            type='transpose2',
            inputs={'X': [input]},
            outputs={'Out': [out],
                     'XShape': [input_shape]},
            attrs={'axis': [1, 0]})
    return out
1113 1114


1115
def cross(x, y, axis=None, name=None):
1116
    """
1117
    Computes the cross product between two tensors along an axis.
1118

1119 1120
    Inputs must have the same shape, and the length of their axes should be equal to 3.
    If `axis` is not given, it defaults to the first axis found with the length 3.
1121

1122
    Args:
1123 1124
        x (Tensor): The first input tensor.
        y (Tensor): The second input tensor.
1125
        axis (int, optional): The axis along which to compute the cross product. It defaults to the first axis found with the length 3.
1126
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.
1127 1128

    Returns:
1129
        Tensor. A Tensor with same data type as `x`.
1130

1131 1132
    Examples:
        .. code-block:: python
1133

1134
            import paddle
1135

Z
Zhou Wei 已提交
1136 1137 1138 1139 1140 1141
            x = paddle.to_tensor([[1.0, 1.0, 1.0],
                                  [2.0, 2.0, 2.0],
                                  [3.0, 3.0, 3.0]])
            y = paddle.to_tensor([[1.0, 1.0, 1.0],
                                  [1.0, 1.0, 1.0],
                                  [1.0, 1.0, 1.0]])
1142

1143 1144 1145 1146 1147 1148 1149 1150 1151
            z1 = paddle.cross(x, y)
            # [[-1. -1. -1.]
            #  [ 2.  2.  2.]
            #  [-1. -1. -1.]]

            z2 = paddle.cross(x, y, axis=1)
            # [[0. 0. 0.]
            #  [0. 0. 0.]
            #  [0. 0. 0.]]
1152
    """
J
Jiabin Yang 已提交
1153 1154 1155 1156 1157 1158 1159 1160
    if in_dygraph_mode():
        return _C_ops.final_state_cross(x, y, axis)
    else:
        if _in_legacy_dygraph():
            if axis is not None:
                return _C_ops.cross(x, y, 'dim', axis)
            else:
                return _C_ops.cross(x, y)
1161
        else:
J
Jiabin Yang 已提交
1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173
            helper = LayerHelper("cross", **locals())
            out = helper.create_variable_for_type_inference(x.dtype)
            attrs = dict()
            attrs['dim'] = axis

            helper.append_op(
                type='cross',
                inputs={'X': x,
                        'Y': y},
                outputs={'Out': out},
                attrs=attrs)
            return out
1174 1175


1176
def cholesky(x, upper=False, name=None):
1177
    r"""
G
Guo Sheng 已提交
1178
    Computes the Cholesky decomposition of one symmetric positive-definite
1179 1180
    matrix or batches of symmetric positive-definite matrice.

G
Guo Sheng 已提交
1181 1182 1183 1184 1185 1186
    If `upper` is `True`, the decomposition has the form :math:`A = U^{T}U` ,
    and the returned matrix :math:`U` is upper-triangular. Otherwise, the
    decomposition has the form  :math:`A = LL^{T}` , and the returned matrix
    :math:`L` is lower-triangular.

    Args:
1187
        x (Tensor): The input tensor. Its shape should be `[*, M, M]`,
G
Guo Sheng 已提交
1188 1189 1190 1191 1192 1193 1194
            where * is zero or more batch dimensions, and matrices on the
            inner-most 2 dimensions all should be symmetric positive-definite.
            Its data type should be float32 or float64.
        upper (bool): The flag indicating whether to return upper or lower
            triangular matrices. Default: False.

    Returns:
1195
        Tensor: A Tensor with same shape and data type as `x`. It represents \
G
Guo Sheng 已提交
1196
            triangular matrices generated by Cholesky decomposition.
1197

G
Guo Sheng 已提交
1198 1199 1200 1201 1202 1203
    Examples:
        .. code-block:: python

            import paddle
            import numpy as np

1204 1205 1206
            a = np.random.rand(3, 3)
            a_t = np.transpose(a, [1, 0])
            x_data = np.matmul(a, a_t) + 1e-03
1207
            x = paddle.to_tensor(x_data)
1208
            out = paddle.linalg.cholesky(x, upper=False)
1209
            print(out)
1210 1211 1212
            # [[1.190523   0.         0.        ]
            #  [0.9906703  0.27676893 0.        ]
            #  [1.25450498 0.05600871 0.06400121]]
G
Guo Sheng 已提交
1213 1214

    """
Z
zhiboniu 已提交
1215
    if paddle.in_dynamic_mode():
W
wanghuancoder 已提交
1216
        return _C_ops.cholesky(x, "upper", upper)
G
Guo Sheng 已提交
1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228
    check_variable_and_dtype(x, 'dtype', ['float32', 'float64'], 'cholesky')
    check_type(upper, 'upper', bool, 'cholesky')
    helper = LayerHelper('cholesky', **locals())
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type='cholesky',
        inputs={'X': [x]},
        outputs={'Out': out},
        attrs={'upper': upper})
    return out


1229 1230 1231 1232
def matrix_rank(x, tol=None, hermitian=False, name=None):
    r"""
    Computes the rank of a matrix.

1233
    The rank of a matrix is the number of singular values that are greater than the specified `tol` threshold when hermitian=False,
1234
    or the number of eigenvalues in absolute value that are greater than the specified `tol` threshold when hermitian=True.
1235 1236

    Args:
1237 1238 1239 1240
        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
1241
            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.
1242 1243
        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
1244
            the lower triangular of the matrix to compute.
1245 1246 1247 1248
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor: Rank of tensor x.
1249

1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265
    Examples:
        .. code-block:: python

            import paddle

            a = paddle.eye(10)
            b = paddle.linalg.matrix_rank(a)
            print(b)
            # b = [10]

            c = paddle.ones(shape=[3, 4, 5, 5])
            d = paddle.linalg.matrix_rank(c, tol=0.01, hermitian=True)
            print(d)
            # d = [[1, 1, 1, 1],
            #      [1, 1, 1, 1],
            #      [1, 1, 1, 1]]
1266

1267 1268
    """

Z
zhiboniu 已提交
1269
    if paddle.in_dynamic_mode():
1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314
        if tol is None:
            tol_tensor = None
            tol_attr = 0.0
            use_default_tol = True
        elif isinstance(tol, Variable):
            if tol.dtype != x.dtype:
                tol_tensor = cast(tol, x.dtype)
            else:
                tol_tensor = tol
            tol_attr = 0.0
            use_default_tol = False
        else:
            tol_tensor = None
            tol_attr = float(tol)
            use_default_tol = False
        return _C_ops.matrix_rank(x, tol_tensor, "tol", tol_attr, 'hermitian',
                                  hermitian, 'use_default_tol', use_default_tol)

    inputs = {}
    attrs = {}
    check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'matrix_rank')
    inputs['X'] = x
    if tol is None:
        attrs['use_default_tol'] = True
    elif isinstance(tol, Variable):
        check_variable_and_dtype(tol, 'tol', ['float32'], 'matrix_rank')
        attrs['use_default_tol'] = False
        if tol.dtype != x.dtype:
            inputs['TolTensor'] = cast(tol, x.dtype)
        else:
            inputs['TolTensor'] = tol
    else:
        check_type(tol, 'tol', float, 'matrix_rank')
        attrs['use_default_tol'] = False
        attrs['tol'] = tol
    check_type(hermitian, 'hermitian', bool, 'matrix_rank')
    attrs['hermitian'] = hermitian

    helper = LayerHelper('matrix_rank', **locals())
    out = helper.create_variable_for_type_inference(dtype='int32')
    helper.append_op(
        type='matrix_rank', inputs=inputs, outputs={'Out': out}, attrs=attrs)
    return out


1315 1316 1317 1318 1319 1320 1321 1322 1323
def bmm(x, y, name=None):
    """
    Applies batched matrix multiplication to two tensors.

    Both of the two input tensors must be three-dementional and share the same batch size.

    if x is a (b, m, k) tensor, y is a (b, k, n) tensor, the output will be a (b, m, n) tensor.

    Args:
Y
yaoxuefeng 已提交
1324 1325
        x (Tensor): The input Tensor.
        y (Tensor): The input Tensor.
1326 1327 1328 1329
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.

    Returns:
Y
yaoxuefeng 已提交
1330
        Tensor: The product Tensor.
1331 1332

    Examples:
S
sunzhongkai588 已提交
1333 1334 1335
        .. code-block:: python

            import paddle
Y
yaoxuefeng 已提交
1336

S
sunzhongkai588 已提交
1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349
            # In imperative mode:
            # size x: (2, 2, 3) and y: (2, 3, 2)
            x = paddle.to_tensor([[[1.0, 1.0, 1.0],
                                [2.0, 2.0, 2.0]],
                                [[3.0, 3.0, 3.0],
                                [4.0, 4.0, 4.0]]])
            y = paddle.to_tensor([[[1.0, 1.0],[2.0, 2.0],[3.0, 3.0]],
                                [[4.0, 4.0],[5.0, 5.0],[6.0, 6.0]]])
            out = paddle.bmm(x, y)
            #output size: (2, 2, 2)
            #output value:
            #[[[6.0, 6.0],[12.0, 12.0]],[[45.0, 45.0],[60.0, 60.0]]]
            out_np = out.numpy()
1350

1351
    """
Y
yaoxuefeng 已提交
1352 1353 1354 1355 1356 1357 1358 1359 1360 1361
    x_shape = x.shape
    y_shape = y.shape
    if not len(x_shape) == len(y_shape) == 3:
        raise ValueError(
            "x and y should be 3-dimensional. But received x's dimention: {}, y's dimention: {}".
            format(x_shape, y_shape))
    if x_shape[2] != y_shape[1]:
        raise ValueError(
            "x's width must be equal with y's height. But received x's shape: {}, y's shape: {}".
            format(x_shape, y_shape))
1362 1363 1364 1365
    if x_shape[0] != y_shape[0]:
        raise ValueError(
            "x's batch (shape[0]) must be equal with y's batch (shape[0]). But received x's shape: {}, y's shape: {}".
            format(x_shape, y_shape))
1366

Z
zhiboniu 已提交
1367
    if paddle.in_dynamic_mode():
W
wanghuancoder 已提交
1368
        return _C_ops.bmm(x, y)
1369 1370

    helper = LayerHelper('bmm', **locals())
1371 1372 1373
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(type='bmm', inputs={'X': x, 'Y': y}, outputs={'Out': out})
    return out
Q
Qi Li 已提交
1374 1375


1376
def histogram(input, bins=100, min=0, max=0, name=None):
Q
Qi Li 已提交
1377
    """
1378
    Computes the histogram of a tensor. The elements are sorted into equal width bins between min and max.
Q
Qi Li 已提交
1379 1380 1381
    If min and max are both zero, the minimum and maximum values of the data are used.

    Args:
1382
        input (Tensor): A Tensor(or LoDTensor) with shape :math:`[N_1, N_2,..., N_k]` . The data type of the input Tensor
Q
Qi Li 已提交
1383 1384 1385 1386 1387 1388
            should be float32, float64, int32, int64.
        bins (int): number of histogram bins
        min (int): lower end of the range (inclusive)
        max (int): upper end of the range (inclusive)

    Returns:
1389
        Tensor: data type is int64, shape is (nbins,).
Q
Qi Li 已提交
1390

1391
    Examples:
Q
Qi Li 已提交
1392
        .. code-block:: python
1393

Q
Qi Li 已提交
1394
            import paddle
1395

1396
            inputs = paddle.to_tensor([1, 2, 1])
1397 1398
            result = paddle.histogram(inputs, bins=4, min=0, max=3)
            print(result) # [0, 2, 1, 0]
Q
Qi Li 已提交
1399
    """
H
hong 已提交
1400 1401 1402 1403
    if in_dygraph_mode():
        return _C_ops.final_state_histogram(input, bins, min, max)

    if _in_legacy_dygraph():
W
wanghuancoder 已提交
1404
        return _C_ops.histogram(input, "bins", bins, "min", min, "max", max)
Q
Qi Li 已提交
1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417

    helper = LayerHelper('histogram', **locals())
    check_variable_and_dtype(
        input, 'X', ['int32', 'int64', 'float32', 'float64'], 'histogram')
    out = helper.create_variable_for_type_inference(VarDesc.VarType.INT64)
    helper.append_op(
        type='histogram',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={'bins': bins,
               'min': min,
               'max': max})
    return out
S
smallv0221 已提交
1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449


def bincount(x, weights=None, minlength=0, name=None):
    """
    Computes frequency of each value in the input tensor. 

    Args:
        x (Tensor): A Tensor with non-negative integer. Should be 1-D tensor.
        weights (Tensor, optional): Weight for each value in the input tensor. Should have the same shape as input. Default is None.
        minlength (int, optional): Minimum number of bins. Should be non-negative integer. Default is 0.
        name(str, optional): The default value is None.  Normally there is no need for user to set this
            property.  For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor: The tensor of frequency.

    Examples:
        .. code-block:: python

            import paddle

            x = paddle.to_tensor([1, 2, 1, 4, 5])
            result1 = paddle.bincount(x)
            print(result1) # [0, 2, 1, 0, 1, 1]

            w = paddle.to_tensor([2.1, 0.4, 0.1, 0.5, 0.5])
            result2 = paddle.bincount(x, weights=w)
            print(result2) # [0., 2.19999981, 0.40000001, 0., 0.50000000, 0.50000000]
    """
    if x.dtype not in [paddle.int32, paddle.int64]:
        raise TypeError("Elements in Input(x) should all be integers")

Z
zhiboniu 已提交
1450
    if paddle.in_dynamic_mode():
S
smallv0221 已提交
1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470
        return _C_ops.bincount(x, weights, "minlength", minlength)

    helper = LayerHelper('bincount', **locals())

    check_variable_and_dtype(x, 'X', ['int32', 'int64'], 'bincount')

    if weights is not None:
        check_variable_and_dtype(weights, 'Weights',
                                 ['int32', 'int64', 'float32', 'float64'],
                                 'bincount')
        out = helper.create_variable_for_type_inference(dtype=weights.dtype)
    else:
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type='bincount',
        inputs={'X': x,
                'Weights': weights},
        outputs={'Out': out},
        attrs={'minlength': minlength})
    return out
1471 1472 1473 1474 1475 1476 1477


def mv(x, vec, name=None):
    """
    Performs a matrix-vector product of the matrix x and the vector vec.

    Args:
F
furnace 已提交
1478
        x (Tensor): A tensor with shape :math:`[M, N]` , The data type of the input Tensor x
1479
            should be one of float32, float64.
F
furnace 已提交
1480
        vec (Tensor): A tensor with shape :math:`[N]` , The data type of the input Tensor x
1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502
            should be one of float32, float64.
        name(str, optional): The default value is None.  Normally there is no need for user to set this
            property.  For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor: The tensor which is producted by x and vec.

    Examples:
        .. code-block:: python

            # x: [M, N], vec: [N]
            # paddle.mv(x, vec)  # out: [M]

            import numpy as np
            import paddle

            x_data = np.array([[2, 1, 3], [3, 0, 1]]).astype("float64")
            x = paddle.to_tensor(x_data)
            vec_data = np.array([3, 5, 1])
            vec = paddle.to_tensor(vec_data).astype("float64")
            out = paddle.mv(x, vec)
    """
J
Jiabin Yang 已提交
1503 1504 1505 1506 1507 1508 1509
    if in_dygraph_mode():
        return _C_ops.final_state_mv(x, vec)
    else:
        if _in_legacy_dygraph():
            out = _C_ops.mv(x, vec)
            return out
        else:
1510

J
Jiabin Yang 已提交
1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534
            def __check_input(x, vec):
                var_names = {'x': x, 'vec': vec}
                for name, val in var_names.items():
                    check_variable_and_dtype(val, name, ['float32', 'float64'],
                                             'mv')
                x_shape = list(x.shape)
                vec_shape = list(vec.shape)
                if len(x_shape) != 2:
                    raise ValueError(
                        "x should be 2-dimensional. But received x's dimention: {}".
                        format(x_shape))
                if len(vec_shape) != 1:
                    raise ValueError(
                        "vec should be 1-dimensional. But received vec's dimention: {}".
                        format(vec_shape))

            __check_input(x, vec)

            helper = LayerHelper('mv', **locals())
            out = helper.create_variable_for_type_inference(dtype=x.dtype)
            helper.append_op(
                type='mv', inputs={'X': x,
                                   'Vec': vec}, outputs={'Out': out})
            return out
1535 1536


1537
def det(x, name=None):
H
huangxu96 已提交
1538 1539 1540 1541 1542 1543 1544 1545
    """
    Calculates determinant value of a square matrix or batches of square matrices.
    Args:
        x (Tensor): input (Tensor): the input matrix of size `(n, n)` or the batch of matrices of size
                    `(*, n, n)` where `*` is one or more batch dimensions.
    Returns:
        y (Tensor):the determinant value of a square matrix or batches of square matrices.

1546
    Examples:
H
huangxu96 已提交
1547 1548 1549 1550 1551 1552
        .. code-block:: python

        import paddle

        x =  paddle.randn([3,3,3])

1553
        A = paddle.linalg.det(x)
H
huangxu96 已提交
1554 1555

        print(A)
1556

H
huangxu96 已提交
1557 1558
        # [ 0.02547996,  2.52317095, -6.15900707])

1559

H
huangxu96 已提交
1560
    """
Z
zhiboniu 已提交
1561
    if paddle.in_dynamic_mode():
1562
        return _C_ops.determinant(x)
H
huangxu96 已提交
1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584

    check_dtype(x.dtype, 'Input', ['float32', 'float64'], 'det')

    input_shape = list(x.shape)
    assert len(input_shape) >= 2,                     \
            "The x must be at least 2-dimensional, "   \
            "but received Input x's dimensional: %s.\n" %  \
            len(input_shape)

    assert (input_shape[-1] == input_shape[-2]),    \
            "Expect squared input," \
            "but received %s by %s matrix.\n" \
            %(input_shape[-2], input_shape[-1]) \

    helper = LayerHelper('determinant', **locals())
    out = helper.create_variable_for_type_inference(dtype=x.dtype)

    helper.append_op(
        type='determinant', inputs={'Input': [x]}, outputs={'Out': [out]})
    return out


1585
def slogdet(x, name=None):
H
huangxu96 已提交
1586 1587 1588
    """
    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)
1589

H
huangxu96 已提交
1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600
    Supports input of float, double

    Note that for matrices that have zero determinant, this returns ``(0, -inf)``
    Args:
        x (Tensor): the batch of matrices of size :math:`(*, n, n)`
            where math:`*` is one or more batch dimensions.

    Returns:
        y (Tensor): A tensor containing the sign of the determinant and the natural logarithm
        of the absolute value of determinant, respectively.

1601
    Examples:
H
huangxu96 已提交
1602 1603 1604 1605 1606 1607
    .. code-block:: python

        import paddle

        x =  paddle.randn([3,3,3])

1608
        A = paddle.linalg.slogdet(x)
H
huangxu96 已提交
1609 1610

        print(A)
1611

H
huangxu96 已提交
1612 1613 1614 1615
        # [[ 1.        ,  1.        , -1.        ],
        # [-0.98610914, -0.43010661, -0.10872950]])

    """
Z
zhiboniu 已提交
1616
    if paddle.in_dynamic_mode():
1617
        return _C_ops.slogdeterminant(x)
H
huangxu96 已提交
1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639

    check_dtype(x.dtype, 'Input', ['float32', 'float64'], 'slogdet')

    input_shape = list(x.shape)
    assert len(input_shape) >= 2,                     \
            "The x must be at least 2-dimensional, "   \
            "but received Input x's dimensional: %s.\n" %  \
            len(input_shape)

    assert (input_shape[-1] == input_shape[-2]),    \
            "Expect squared input," \
            "but received %s by %s matrix.\n" \
            %(input_shape[-2], input_shape[-1]) \

    helper = LayerHelper('slogdeterminant', **locals())
    out = helper.create_variable_for_type_inference(dtype=x.dtype)

    helper.append_op(
        type='slogdeterminant', inputs={'Input': [x]}, outputs={'Out': [out]})
    return out


1640 1641
def svd(x, full_matrices=False, name=None):
    r"""
1642 1643 1644 1645 1646
    Computes the singular value decomposition of one matrix or a batch of regular matrices.

    Let :math:`X` be the input matrix or a batch of input matrices, the output should satisfies:

    .. math::
1647 1648
        X = U * diag(S) * VT

1649 1650
    Args:
        x (Tensor): The input tensor. Its shape should be `[..., N, M]`,
1651
            where `...` is zero or more batch dimensions. N and M can be arbitraty
1652 1653 1654 1655
            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,
1656
            which means shape of U is `[..., N, N]`, shape of V is `[..., M, M]`. K = min(M, N).
1657
            If full_matrices = False, svd op will use a economic method to store U and V.
1658
            which means shape of U is `[..., N, K]`, shape of V is `[..., M, K]`. K = min(M, N).
1659
        name (str, optional): Name for the operation (optional, default is None).
1660
            For more information, please refer to :ref:`api_guide_Name`.
1661 1662

    Returns:
1663
        Tuple of 3 tensors: (U, S, VH). VH is the conjugate transpose of V. S is the singlar value vectors of matrics with shape `[..., K]`
1664

1665 1666 1667 1668
    Examples:
        .. code-block:: python

            import paddle
1669 1670 1671

            x = paddle.to_tensor([[1.0, 2.0], [1.0, 3.0], [4.0, 6.0]]).astype('float64')
            x = x.reshape([3, 2])
1672
            u, s, vh = paddle.linalg.svd(x)
1673 1674 1675 1676 1677
            print (u)
            #U = [[ 0.27364809, -0.21695147  ],
            #      [ 0.37892198, -0.87112408 ],
            #      [ 0.8840446 ,  0.44053933 ]]

1678
            print (s)
1679
            #S = [8.14753743, 0.78589688]
1680
            print (vh)
1681 1682
            #VT= [[ 0.51411221,  0.85772294],
            #     [ 0.85772294, -0.51411221]]
1683

1684
            # one can verify : U * S * VT == X
1685
            #                  U * UH == I
1686
            #                  V * VH == I
1687 1688
    """

Z
zhiboniu 已提交
1689
    if paddle.in_dynamic_mode():
1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704
        return _C_ops.svd(x, 'full_matrices', full_matrices)
    check_variable_and_dtype(x, 'dtype', ['float32', 'float64'], 'svd')
    check_type(full_matrices, 'full_matrices', bool, 'svd')
    helper = LayerHelper('svd', **locals())
    u = helper.create_variable_for_type_inference(dtype=x.dtype)
    vh = helper.create_variable_for_type_inference(dtype=x.dtype)
    s = helper.create_variable_for_type_inference(dtype=x.dtype)
    attrs = dict()
    attrs['full_matrices'] = full_matrices
    helper.append_op(
        type='svd',
        inputs={'X': [x]},
        outputs={'U': u,
                 'VH': vh,
                 'S': s},
X
xiongkun 已提交
1705
        attrs=attrs, )
1706 1707 1708
    return u, s, vh


1709 1710 1711
def matrix_power(x, n, name=None):
    r"""
    Computes the n-th power of a square matrix or a batch of square matrices.
1712

1713 1714 1715 1716 1717
    Let :math:`X` be a sqaure matrix or a batch of square matrices, :math:`n` be
    an exponent, the equation should be:

    .. math::
        Out = X ^ {n}
1718

1719 1720 1721 1722
    Specifically,

    - If `n > 0`, it returns the matrix or a batch of matrices raised to the power
    of `n`.
1723

1724 1725 1726 1727 1728 1729 1730 1731 1732 1733
    - If `n = 0`, it returns the identity matrix or a batch of identity matrices.

    - If `n < 0`, it returns the inverse of each matrix (if invertible) raised to
    the power of `abs(n)`.

    Args:
        x (Tensor): A square matrix or a batch of square matrices to be raised
            to power `n`. Its shape should be `[*, M, M]`, where `*` is zero or
            more batch dimensions. Its data type should be float32 or float64.
        n (int): The exponent. It can be any positive, negative integer or zero.
1734
        name (str, optional): Name for the operation (optional, default is None).
1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748
            For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor: The n-th power of the matrix (or the batch of matrices) `x`. Its
            data type should be the same as that of `x`.

    Examples:
        .. code-block:: python

            import paddle

            x = paddle.to_tensor([[1, 2, 3],
                                  [1, 4, 9],
                                  [1, 8, 27]], dtype='float64')
1749
            print(paddle.linalg.matrix_power(x, 2))
1750 1751 1752 1753
            # [[6.  , 34. , 102.],
            #  [14. , 90. , 282.],
            #  [36. , 250., 804.]]

1754
            print(paddle.linalg.matrix_power(x, 0))
1755 1756 1757 1758
            # [[1., 0., 0.],
            #  [0., 1., 0.],
            #  [0., 0., 1.]]

1759
            print(paddle.linalg.matrix_power(x, -2))
1760 1761 1762 1763
            # [[ 12.91666667, -12.75000000,  2.83333333 ],
            #  [-7.66666667 ,  8.         , -1.83333333 ],
            #  [ 1.80555556 , -1.91666667 ,  0.44444444 ]]
    """
Z
zhiboniu 已提交
1764
    if paddle.in_dynamic_mode():
1765
        return _C_ops.matrix_power(x, "n", n)
1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776

    check_variable_and_dtype(x, 'dtype', ['float32', 'float64'], 'matrix_power')
    check_type(n, 'n', int, 'matrix_power')
    helper = LayerHelper('matrix_power', **locals())
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type='matrix_power',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'n': n})
    return out
1777 1778


1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820
def qr(x, mode="reduced", name=None):
    r"""
    Computes the QR decomposition of one matrix or batches of matrice (backward is unsupported now).

    Args:
        x (Tensor): The input tensor. Its shape should be `[..., M, N]`,
            where ... is zero or more batch dimensions. M and N can be arbitrary
            positive number. The data type of x should be float32 or float64. 
        mode (str, optional): A flag to control the behavior of qr, the default is "reduced". 
            Suppose x's shape is `[..., M, N]` and denoting `K = min(M, N)`:
            If mode = "reduced", qr op will return reduced Q and R matrices, 
            which means Q's shape is `[..., M, K]` and R's shape is `[..., K, N]`.
            If mode = "complete", qr op will return complete Q and R matrices, 
            which means Q's shape is `[..., M, M]` and R's shape is `[..., M, N]`.
            If mode = "r", qr op will only return reduced R matrix, which means
            R's shape is `[..., K, N]`.
        name (str, optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.
            
    Returns:
        If mode = "reduced" or mode = "complete", qr will return a two tensor-tuple, which represents Q and R. 
        If mode = "r", qr will return a tensor which represents R.
        
    Examples:            
        .. code-block:: python

            import paddle 

            x = paddle.to_tensor([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]]).astype('float64')
            q, r = paddle.linalg.qr(x)
            print (q)
            print (r)

            # Q = [[-0.16903085,  0.89708523],
            #      [-0.50709255,  0.27602622],
            #      [-0.84515425, -0.34503278]])

            # R = [[-5.91607978, -7.43735744],
            #      [ 0.        ,  0.82807867]])
            
            # one can verify : X = Q * R ;     
    """
Z
zhiboniu 已提交
1821
    if paddle.in_dynamic_mode():
1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842
        q, r = _C_ops.qr(x, 'mode', mode)
        if mode == "r":
            return r
        else:
            return q, r
    check_variable_and_dtype(x, 'dtype', ['float32', 'float64'], 'qr')
    check_type(mode, 'mode', str, 'qr')
    helper = LayerHelper('qr', **locals())
    q = helper.create_variable_for_type_inference(dtype=x.dtype)
    r = helper.create_variable_for_type_inference(dtype=x.dtype)
    attrs = dict()
    attrs['mode'] = mode
    helper.append_op(
        type='qr', inputs={'X': [x]}, outputs={'Q': q,
                                               'R': r}, attrs=attrs)
    if mode == "r":
        return r
    else:
        return q, r


1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919
def lu(x, pivot=True, get_infos=False, name=None):
    r"""
    Computes the LU factorization of an N-D(N>=2) matrix x. 

    Returns the LU factorization(inplace x) and Pivots. low triangular matrix L and 
    upper triangular matrix U are combined to a single LU matrix.

    Pivoting is done if pivot is set to True.
    P mat can be get by pivots:
    # ones = eye(rows) #eye matrix of rank rows
    # for i in range(cols):
    #     swap(ones[i], ones[pivots[i]])
    # return ones

    Args:

        X (Tensor): the tensor to factor of N-dimensions(N>=2).

        pivot (bool, optional): controls whether pivoting is done. Default: True.

        get_infos (bool, optional): if set to True, returns an info IntTensor. Default: False.

        name (str, optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.
            
    Returns:
        factorization (Tensor): LU matrix, the factorization of input X.

        pivots (IntTensor): the pivots of size(∗(N-2), min(m,n)). `pivots` stores all the 
                    intermediate transpositions of rows. The final permutation `perm` could be 
                    reconstructed by this, details refer to upper example.

        infos (IntTensor, optional): if `get_infos` is `True`, this is a tensor of size (∗(N-2)) 
                    where non-zero values indicate whether factorization for the matrix or each minibatch 
                    has succeeded or failed.

        
    Examples:            
        .. code-block:: python

            import paddle 

            x = paddle.to_tensor([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]]).astype('float64')
            lu,p,info = paddle.linalg.lu(x, get_infos=True)

            # >>> lu:
            # Tensor(shape=[3, 2], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
            #    [[5.        , 6.        ],
            #        [0.20000000, 0.80000000],
            #        [0.60000000, 0.50000000]])
            # >>> p
            # Tensor(shape=[2], dtype=int32, place=CUDAPlace(0), stop_gradient=True,
            #    [3, 3])
            # >>> info
            # Tensor(shape=[], dtype=int32, place=CUDAPlace(0), stop_gradient=True,
            #    0)
            
            P,L,U = paddle.linalg.lu_unpack(lu,p)

            # >>> P
            # (Tensor(shape=[3, 3], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
            # [[0., 1., 0.],
            # [0., 0., 1.],
            # [1., 0., 0.]]), 
            # >>> L
            # Tensor(shape=[3, 2], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
            # [[1.        , 0.        ],
            # [0.20000000, 1.        ],
            # [0.60000000, 0.50000000]]), 
            # >>> U
            # Tensor(shape=[2, 2], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
            # [[5.        , 6.        ],
            # [0.        , 0.80000000]]))
            

            # one can verify : X = P @ L @ U ;     
    """
Z
zhiboniu 已提交
1920
    if paddle.in_dynamic_mode():
1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
        LU, Piv, Info = _C_ops.lu(x, 'pivots', pivot)
        if get_infos:
            return LU, Piv, Info
        else:
            return LU, Piv
    check_variable_and_dtype(x, 'dtype', ['float32', 'float64'], 'lu')
    helper = LayerHelper('lu', **locals())
    lu = helper.create_variable_for_type_inference(dtype=x.dtype)
    p = helper.create_variable_for_type_inference(dtype='int')
    info = helper.create_variable_for_type_inference(dtype='int')
    attrs = dict()
    attrs['pivots'] = pivot
    helper.append_op(
        type='lu',
        inputs={'X': x},
        outputs={'Out': lu,
                 'Pivots': p,
                 'Infos': info},
        attrs=attrs)
    if get_infos:
        return lu, p, info
    else:
        return lu, p


def lu_unpack(x, y, unpack_ludata=True, unpack_pivots=True, name=None):
    r"""
    Unpack L U and P to single matrix tensor . 
    unpack L and U matrix from LU, unpack permutation matrix P from Pivtos .

    P mat can be get by pivots:
    # ones = eye(rows) #eye matrix of rank rows
    # for i in range(cols):
    #     swap(ones[i], ones[pivots[i]])


    Args:
        x (Tensor): The LU tensor get from paddle.linalg.lu, which is combined by L and U.

        y (Tensor): Pivots get from paddle.linalg.lu.

        unpack_ludata (bool,optional): whether to unpack L and U from x. Default: True.

        unpack_pivots (bool, optional): whether to unpack permutation matrix P from Pivtos. Default: True.

        name (str, optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.
            
    Returns:
        P (Tensor): Permutation matrix P of lu factorization.

        L (Tensor): The lower triangular matrix tensor of lu factorization.

        U (Tensor): The upper triangular matrix tensor of lu factorization.

        
    Examples:            
        .. code-block:: python

            import paddle 

            x = paddle.to_tensor([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]]).astype('float64')
            lu,p,info = paddle.linalg.lu(x, get_infos=True)

            # >>> lu:
            # Tensor(shape=[3, 2], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
            #    [[5.        , 6.        ],
            #        [0.20000000, 0.80000000],
            #        [0.60000000, 0.50000000]])
            # >>> p
            # Tensor(shape=[2], dtype=int32, place=CUDAPlace(0), stop_gradient=True,
            #    [3, 3])
            # >>> info
            # Tensor(shape=[], dtype=int32, place=CUDAPlace(0), stop_gradient=True,
            #    0)
            
            P,L,U = paddle.linalg.lu_unpack(lu,p)

            # >>> P
            # (Tensor(shape=[3, 3], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
            # [[0., 1., 0.],
            # [0., 0., 1.],
            # [1., 0., 0.]]), 
            # >>> L
            # Tensor(shape=[3, 2], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
            # [[1.        , 0.        ],
            # [0.20000000, 1.        ],
            # [0.60000000, 0.50000000]]), 
            # >>> U
            # Tensor(shape=[2, 2], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
            # [[5.        , 6.        ],
            # [0.        , 0.80000000]]))

            # one can verify : X = P @ L @ U ;   
    """

Z
zhiboniu 已提交
2017
    if paddle.in_dynamic_mode():
2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041
        P, L, U = _C_ops.lu_unpack(x, y, 'unpack_ludata', unpack_ludata,
                                   'unpack_pivots', unpack_pivots)
        return P, L, U

    check_variable_and_dtype(x, 'dtype', ['float32', 'float64'], 'lu_unpack')
    helper = LayerHelper('lu_unpack', **locals())
    p = helper.create_variable_for_type_inference(dtype=x.dtype)
    l = helper.create_variable_for_type_inference(dtype=x.dtype)
    u = helper.create_variable_for_type_inference(dtype=x.dtype)

    attrs = dict()
    attrs['unpack_ludata'] = unpack_ludata
    attrs['unpack_pivots'] = unpack_pivots
    helper.append_op(
        type='lu_unpack',
        inputs={'X': x,
                'Pivots': y},
        outputs={'Pmat': p,
                 'L': l,
                 'U': u},
        attrs=attrs)
    return p, l, u


L
Lijunhui 已提交
2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089
def eig(x, name=None):
    """
    This API performs the eigenvalue decomposition of a square matrix or a batch of square matrices.

    .. note::
        If the matrix is a Hermitian or a real symmetric matrix, please use :ref:`paddle.linalg.eigh` instead, which is much faster.
        If only eigenvalues is needed, please use :ref:`paddle.linalg.eigvals` instead.
        If the matrix is of any shape, please use :ref:`paddle.linalg.svd`.
        This API is only supported on CPU device.
        The output datatype is always complex for both real and complex input.

    Args:
        x (Tensor): A tensor with shape math:`[*, N, N]`, The data type of the x should be one of ``float32``,
            ``float64``, ``compplex64`` or ``complex128``.
        name (str, optional): The default value is `None`. Normally there is no need for user to set 
            this property. For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Eigenvalues(Tensors): A tensor with shape math:`[*, N]` refers to the eigen values.
        Eigenvectors(Tensors): A tensor with shape math:`[*, N, N]` refers to the eigen vectors.

    Examples:
        .. code-block:: python

            import paddle
            import numpy as np

            paddle.device.set_device("cpu")

            x_data = np.array([[1.6707249, 7.2249975, 6.5045543],
                               [9.956216,  8.749598,  6.066444 ],
                               [4.4251957, 1.7983172, 0.370647 ]]).astype("float32")
            x = paddle.to_tensor(x_data)
            w, v = paddle.linalg.eig(x)
            print(w)
            # Tensor(shape=[3, 3], dtype=complex128, place=CPUPlace, stop_gradient=False,
            #       [[(-0.5061363550800655+0j) , (-0.7971760990842826+0j) ,
            #         (0.18518077798279986+0j)],
            #        [(-0.8308237755993192+0j) ,  (0.3463813401919749+0j) ,
            #         (-0.6837005269141947+0j) ],
            #        [(-0.23142567697893396+0j),  (0.4944999840400175+0j) ,
            #         (0.7058765252952796+0j) ]])

            print(v)
            # Tensor(shape=[3], dtype=complex128, place=CPUPlace, stop_gradient=False,
            #       [ (16.50471283351188+0j)  , (-5.5034820550763515+0j) ,
            #         (-0.21026087843552282+0j)])
    """
Z
zhiboniu 已提交
2090
    if paddle.in_dynamic_mode():
L
Lijunhui 已提交
2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107
        w, v = _C_ops.eig(x)
        return w, v

    check_variable_and_dtype(
        x, 'X', ['float32', 'float64', 'complex64', 'complex128'], 'eig')
    helper = LayerHelper('eig', **locals())

    w = helper.create_variable_for_type_inference(x.dtype)
    v = helper.create_variable_for_type_inference(x.dtype)

    inputs = {'X': x}
    outputs = {'Eigenvalues': w, 'Eigenvectors': v}
    helper.append_op(type='eig', inputs=inputs, outputs=outputs)

    return w, v


2108 2109 2110
def eigvals(x, name=None):
    """
    Compute the eigenvalues of one or more general matrices.
2111 2112 2113

    Warning:
        The gradient kernel of this operator does not yet developed.
2114 2115 2116 2117
        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.
2118
            Its shape should be `[*, M, M]`, where `*` is zero or more batch dimensions.
2119
            Its data type should be float32, float64, complex64, or complex128.
2120
        name (str, optional): Name for the operation (optional, default is None).
2121
            For more information, please refer to :ref:`api_guide_Name`.
2122
            
2123
    Returns:
2124
        Tensor: A tensor containing the unsorted eigenvalues which has the same batch dimensions with `x`.
2125 2126 2127 2128 2129 2130
            The eigenvalues are complex-valued even when `x` is real.

    Examples:
        .. code-block:: python

            import paddle
2131

2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144
            paddle.set_device("cpu")
            paddle.seed(1234)

            x = paddle.rand(shape=[3, 3], dtype='float64')
            # [[0.02773777, 0.93004224, 0.06911496],
            #  [0.24831591, 0.45733623, 0.07717843],
            #  [0.48016702, 0.14235102, 0.42620817]])

            print(paddle.linalg.eigvals(x))
            # [(-0.27078833542132674+0j), (0.29962280156230725+0j), (0.8824477020120244+0j)] #complex128
    """

    check_variable_and_dtype(x, 'dtype',
2145 2146
                             ['float32', 'float64', 'complex64',
                              'complex128'], 'eigvals')
2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158

    x_shape = list(x.shape)
    if len(x_shape) < 2:
        raise ValueError(
            "The dimension of Input(x) should be at least 2, but received x's dimention = {}, x's shape = {}".
            format(len(x_shape), x_shape))

    if x_shape[-1] != x_shape[-2]:
        raise ValueError(
            "The last two dimensions of Input(x) should be equal, but received x's shape = {}".
            format(x_shape))

Z
zhiboniu 已提交
2159
    if paddle.in_dynamic_mode():
2160 2161 2162 2163 2164 2165 2166 2167
        return _C_ops.eigvals(x)

    helper = LayerHelper('eigvals', **locals())
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(type='eigvals', inputs={'X': x}, outputs={'Out': out})
    return out


2168 2169 2170 2171
def multi_dot(x, name=None):
    """
    Multi_dot is an operator that calculates multiple matrix multiplications.

2172
    Supports inputs of float16(only GPU support), float32 and float64 dtypes. This function does not
2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213
    support batched inputs.

    The input tensor in [x] must be 2-D except for the first and last can be 1-D.
    If the first tensor is a 1-D vector of shape(n, ) it is treated as row vector
    of shape(1, n), similarly if the last tensor is a 1D vector of shape(n, ), it
    is treated as a column vector of shape(n, 1).

    If the first and last tensor are 2-D matrix, then the output is also 2-D matrix,
    otherwise the output is a 1-D vector.

    Multi_dot will select the lowest cost multiplication order for calculation. The
    cost of multiplying two matrices with shapes (a, b) and (b, c) is a * b * c.
    Given matrices A, B, C with shapes (20, 5), (5, 100), (100, 10) respectively,
    we can calculate the cost of different multiplication orders as follows:
    - Cost((AB)C) = 20x5x100 + 20x100x10 = 30000
    - Cost(A(BC)) = 5x100x10 + 20x5x10 = 6000

    In this case, multiplying B and C first, then multiply A, which is 5 times faster
    than sequential calculation.

    Args:
        x ([Tensor]): The input tensors which is a list Tensor.
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.

    Returns:
        Tensor: The output Tensor.


    Examples:

    .. code-block:: python

        import paddle
        import numpy as np

        # A * B
        A_data = np.random.random([3, 4]).astype(np.float32)
        B_data = np.random.random([4, 5]).astype(np.float32)
        A = paddle.to_tensor(A_data)
        B = paddle.to_tensor(B_data)
2214
        out = paddle.linalg.multi_dot([A, B])
2215 2216 2217 2218 2219 2220 2221 2222 2223 2224
        print(out.numpy().shape)
        # [3, 5]

        # A * B * C
        A_data = np.random.random([10, 5]).astype(np.float32)
        B_data = np.random.random([5, 8]).astype(np.float32)
        C_data = np.random.random([8, 7]).astype(np.float32)
        A = paddle.to_tensor(A_data)
        B = paddle.to_tensor(B_data)
        C = paddle.to_tensor(C_data)
2225
        out = paddle.linalg.multi_dot([A, B, C])
2226 2227 2228 2229
        print(out.numpy().shape)
        # [10, 7]

    """
Z
zhiboniu 已提交
2230
    if paddle.in_dynamic_mode():
2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245
        return _C_ops.multi_dot(x)

    check_type(x, 'x', (list, tuple), 'multi_dot')
    for id, item in enumerate(x):
        check_variable_and_dtype(item, 'x[' + str(id) + ']',
                                 ['float16', 'float32', 'float64'], 'multi_dot')
        if item.dtype != x[0].dtype:
            raise TypeError(
                "All the Tensors in the input must have the same data type.")

    helper = LayerHelper('multi_dot', **locals())
    dtype = helper.input_dtype(input_param_name='x')
    out = helper.create_variable_for_type_inference(dtype)
    helper.append_op(type='multi_dot', inputs={"X": x}, outputs={"Out": out})
    return out
2246 2247 2248 2249


def eigh(x, UPLO='L', name=None):
    """
2250
    Compute the eigenvalues and eigenvectors of a
2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273
    complex Hermitian (conjugate symmetric) or a real symmetric matrix.

    Args:
        x (Tensor): A tensor with shape :math:`[*, N, N]` , The data type of the input Tensor x
            should be one of float32, float64, complex64, complex128.
        UPLO(str, optional): (string, default 'L'), 'L' represents the lower triangular matrix,
                        "'U' represents the upper triangular matrix.".
        name(str, optional): The default value is None.  Normally there is no need for user to set this
            property.  For more information, please refer to :ref:`api_guide_Name`.

    Returns:

        out_value(Tensor):  A Tensor with shape [*, N] and data type of float32 and float64. The eigenvalues of eigh op.
        out_vector(Tensor): A Tensor with shape [*, N, N] and data type of float32,float64,complex64 and complex128. The eigenvectors of eigh op.

    Examples:
        .. code-block:: python

            import numpy as np
            import paddle

            x_data = np.array([[1, -2j], [2j, 5]])
            x = paddle.to_tensor(x_data)
2274
            out_value, out_vector = paddle.linalg.eigh(x, UPLO='L')
2275 2276 2277 2278 2279 2280 2281
            print(out_value)
            #[0.17157288, 5.82842712]
            print(out_vector)
            #[(-0.9238795325112867+0j), (-0.3826834323650898+0j)],
            #[ 0.3826834323650898j    , -0.9238795325112867j    ]]

    """
Z
zhiboniu 已提交
2282
    if paddle.in_dynamic_mode():
2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294
        return _C_ops.eigh(x, 'UPLO', UPLO)

    def __check_input(x, UPLO):
        x_shape = list(x.shape)
        if len(x.shape) < 2:
            raise ValueError(
                "Input(input) only support >=2 tensor, but received "
                "length of Input(input) is %s." % len(x.shape))
        if x_shape[-1] != x_shape[-2]:
            raise ValueError(
                "The input matrix must be batches of square matrices. But received x's dimention: {}".
                format(x_shape))
2295
        if UPLO != 'L' and UPLO != 'U':
2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314
            raise ValueError(
                "UPLO must be L or U. But received UPLO is: {}".format(UPLO))

    __check_input(x, UPLO)

    helper = LayerHelper('eigh', **locals())
    check_variable_and_dtype(
        x, 'dtype', ['float32', 'float64', 'complex64', 'complex128'], 'eigh')

    out_value = helper.create_variable_for_type_inference(dtype=x.dtype)
    out_vector = helper.create_variable_for_type_inference(dtype=x.dtype)

    helper.append_op(
        type='eigh',
        inputs={'X': x},
        outputs={'Eigenvalues': out_value,
                 'Eigenvectors': out_vector},
        attrs={'UPLO': UPLO})
    return out_value, out_vector
A
andyjpaddle 已提交
2315 2316 2317 2318


def pinv(x, rcond=1e-15, hermitian=False, name=None):
    r"""
2319
    Calculate pseudo inverse via SVD(singular value decomposition)
A
andyjpaddle 已提交
2320 2321 2322 2323 2324 2325 2326 2327 2328 2329
    of one matrix or batches of regular matrix.

    .. math::

        if hermitian == False:
            x = u * s * vt  (SVD)
            out = v * 1/s * ut
        else:
            x = u * s * ut  (eigh)
            out = u * 1/s * u.conj().transpose(-2,-1)
2330

A
andyjpaddle 已提交
2331 2332 2333
    If x is hermitian or symmetric matrix, svd will be replaced with eigh.

    Args:
2334 2335 2336
        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
A
andyjpaddle 已提交
2337 2338 2339 2340
            float32 or float64 or complex64 or complex128. When data
            type is complex64 or cpmplex128, hermitian should be set
            True.

2341 2342 2343 2344
        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
A
andyjpaddle 已提交
2345
            if complex or symmetric if real. Default: False.
2346 2347

        name(str|None): A name for this layer(optional). If set None,
A
andyjpaddle 已提交
2348
            the layer will be named automatically.
2349

A
andyjpaddle 已提交
2350
    Returns:
2351
        Tensor: The tensor with same data type with x. it represents
A
andyjpaddle 已提交
2352
        pseudo inverse of x. Its shape should be (*, n, m).
2353

A
andyjpaddle 已提交
2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380
    Examples:
        .. code-block:: python

            import paddle

            x = paddle.arange(15).reshape((3, 5)).astype('float64')
            input = paddle.to_tensor(x)
            out = paddle.linalg.pinv(input)
            print(input)
            print(out)

            # input:
            # [[0. , 1. , 2. , 3. , 4. ],
            # [5. , 6. , 7. , 8. , 9. ],
            # [10., 11., 12., 13., 14.]]

            # out:
            # [[-0.22666667, -0.06666667,  0.09333333],
            # [-0.12333333, -0.03333333,  0.05666667],
            # [-0.02000000,  0.00000000,  0.02000000],
            # [ 0.08333333,  0.03333333, -0.01666667],
            # [ 0.18666667,  0.06666667, -0.05333333]]

            # one can verify : x * out * x = x ;
            # or              out * x * out = x ;
    """

Z
zhiboniu 已提交
2381
    if paddle.in_dynamic_mode():
A
andyjpaddle 已提交
2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512
        if not hermitian:
            # combine svd and matmul op
            u, s, vt = _C_ops.svd(x, 'full_matrices', False)
            max_singular_val = _C_ops.reduce_max(s, 'dim', [-1], 'keep_dim', True, \
                'reduce_all', False)
            rcond = paddle.to_tensor(rcond, dtype=x.dtype)
            cutoff = rcond * max_singular_val
            y = float('inf')
            y = paddle.to_tensor(y, dtype=x.dtype)

            condition = s > cutoff
            cond_int = layers.cast(condition, s.dtype)
            cond_not_int = layers.cast(layers.logical_not(condition), s.dtype)
            out1 = layers.elementwise_mul(1 / s, cond_int)
            out2 = layers.elementwise_mul(1 / y, cond_not_int)
            singular = layers.elementwise_add(out1, out2)
            st, _ = _C_ops.unsqueeze2(singular, 'axes', [-2])

            dims = list(range(len(vt.shape)))
            perm = dims[:-2] + [dims[-1]] + [dims[-2]]
            v, _ = _C_ops.transpose2(vt, 'axis', perm)

            out_1 = v * st
            out_2 = _C_ops.matmul_v2(out_1, u, 'trans_x', False, 'trans_y',
                                     True)
            return out_2
        else:
            # combine eigh and matmul op
            s, u = _C_ops.eigh(x, 'UPLO', 'L')
            s_abs = paddle.abs(s)
            max_singular_val = _C_ops.reduce_max(s_abs, 'dim', [-1], 'keep_dim', True, \
                'reduce_all', False)
            rcond = paddle.to_tensor(rcond, dtype=s.dtype)
            cutoff = rcond * max_singular_val
            y = float('inf')
            y = paddle.to_tensor(y, dtype=s.dtype)

            condition = s_abs > cutoff
            cond_int = layers.cast(condition, s.dtype)
            cond_not_int = layers.cast(layers.logical_not(condition), s.dtype)
            out1 = layers.elementwise_mul(1 / s, cond_int)
            out2 = layers.elementwise_mul(1 / y, cond_not_int)
            singular = layers.elementwise_add(out1, out2)
            st, _ = _C_ops.unsqueeze2(singular, 'axes', [-2])

            out_1 = u * st
            u_conj = _C_ops.conj(u)
            out_2 = _C_ops.matmul_v2(out_1, u_conj, 'trans_x', False, 'trans_y',
                                     True)
            return out_2
    else:
        if not hermitian:
            helper = LayerHelper('pinv', **locals())
            dtype = x.dtype
            check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'pinv')

            u = helper.create_variable_for_type_inference(dtype)
            s = helper.create_variable_for_type_inference(dtype)
            vt = helper.create_variable_for_type_inference(dtype)
            helper.append_op(
                type='svd',
                inputs={'X': [x]},
                outputs={'U': u,
                         'VH': vt,
                         'S': s},
                attrs={'full_matrices': False}, )

            max_singular_val = helper.create_variable_for_type_inference(dtype)
            helper.append_op(
                type='reduce_max',
                inputs={'X': s},
                outputs={'Out': max_singular_val},
                attrs={'dim': [-1],
                       'keep_dim': True,
                       'reduce_all': False})

            rcond = layers.fill_constant(shape=[1], value=rcond, dtype=dtype)
            cutoff = rcond * max_singular_val
            y = float('inf')
            y = layers.fill_constant(shape=[1], value=y, dtype=dtype)

            condition = s > cutoff
            cond_int = layers.cast(condition, dtype)
            cond_not_int = layers.cast(layers.logical_not(condition), dtype)
            out1 = layers.elementwise_mul(1 / s, cond_int)
            out2 = layers.elementwise_mul(1 / y, cond_not_int)
            singular = layers.elementwise_add(out1, out2)

            st = helper.create_variable_for_type_inference(dtype=dtype)
            st_shape = helper.create_variable_for_type_inference(dtype=dtype)
            helper.append_op(
                type='unsqueeze2',
                inputs={'X': singular},
                attrs={'axes': [-2]},
                outputs={'Out': st,
                         'XShape': st_shape})

            dims = list(range(len(vt.shape)))
            perm = dims[:-2] + [dims[-1]] + [dims[-2]]
            v = helper.create_variable_for_type_inference(dtype)
            v_shape = helper.create_variable_for_type_inference(dtype)
            helper.append_op(
                type='transpose2',
                inputs={'X': [vt]},
                outputs={'Out': [v],
                         'XShape': [v_shape]},
                attrs={'axis': perm})

            out_1 = helper.create_variable_for_type_inference(dtype)
            helper.append_op(
                type='elementwise_mul',
                inputs={'X': v,
                        'Y': st},
                outputs={'Out': out_1},
                attrs={'axis': -1,
                       'use_mkldnn': False})
            out_1 = helper.append_activation(out_1)

            out_2 = helper.create_variable_for_type_inference(dtype)
            helper.append_op(
                type='matmul_v2',
                inputs={'X': out_1,
                        'Y': u},
                outputs={'Out': out_2},
                attrs={'trans_x': False,
                       'trans_y': True}, )
            return out_2
        else:
            helper = LayerHelper('pinv', **locals())
            dtype = x.dtype
            check_variable_and_dtype(
2513 2514
                x, 'dtype', ['float32', 'float64', 'complex64',
                             'complex128'], 'pinv')
A
andyjpaddle 已提交
2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586

            if dtype == paddle.complex128:
                s_type = 'float64'
            elif dtype == paddle.complex64:
                s_type = 'float32'
            else:
                s_type = dtype

            u = helper.create_variable_for_type_inference(dtype)
            s = helper.create_variable_for_type_inference(s_type)
            helper.append_op(
                type='eigh',
                inputs={'X': x},
                outputs={'Eigenvalues': s,
                         'Eigenvectors': u},
                attrs={'UPLO': 'L'})
            s_abs = helper.create_variable_for_type_inference(s_type)
            helper.append_op(
                type='abs', inputs={'X': s}, outputs={'Out': s_abs})
            max_singular_val = helper.create_variable_for_type_inference(s_type)
            helper.append_op(
                type='reduce_max',
                inputs={'X': s_abs},
                outputs={'Out': max_singular_val},
                attrs={'dim': [-1],
                       'keep_dim': True,
                       'reduce_all': False})

            rcond = layers.fill_constant(shape=[1], value=rcond, dtype=s_type)
            cutoff = rcond * max_singular_val
            y = float('inf')
            y = layers.fill_constant(shape=[1], value=y, dtype=s_type)

            condition = s_abs > cutoff
            cond_int = layers.cast(condition, s_type)
            cond_not_int = layers.cast(layers.logical_not(condition), s_type)
            out1 = layers.elementwise_mul(1 / s, cond_int)
            out2 = layers.elementwise_mul(1 / y, cond_not_int)
            singular = layers.elementwise_add(out1, out2)

            st = helper.create_variable_for_type_inference(dtype=s_type)
            st_shape = helper.create_variable_for_type_inference(dtype=s_type)
            helper.append_op(
                type='unsqueeze2',
                inputs={'X': singular},
                attrs={'axes': [-2]},
                outputs={'Out': st,
                         'XShape': st_shape})

            out_1 = helper.create_variable_for_type_inference(dtype)
            helper.append_op(
                type='elementwise_mul',
                inputs={'X': u,
                        'Y': st},
                outputs={'Out': out_1},
                attrs={'axis': -1,
                       'use_mkldnn': False})
            out_1 = helper.append_activation(out_1)

            u_conj = helper.create_variable_for_type_inference(dtype)
            helper.append_op(
                type='conj', inputs={'X': u}, outputs={'Out': [u_conj]})

            out_2 = helper.create_variable_for_type_inference(dtype)
            helper.append_op(
                type='matmul_v2',
                inputs={'X': out_1,
                        'Y': u_conj},
                outputs={'Out': out_2},
                attrs={'trans_x': False,
                       'trans_y': True}, )
            return out_2
W
Weilong Wu 已提交
2587 2588 2589 2590 2591 2592 2593


def solve(x, y, name=None):
    r"""
    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:
2594

W
Weilong Wu 已提交
2595 2596 2597 2598
    .. math::
        Out = X^-1 * Y
    Specifically,
    - This system of linear equations has one solution if and only if input 'X' is invertible.
2599

W
Weilong Wu 已提交
2600 2601 2602 2603 2604
    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.
2605
        name(str, optional): Name for the operation (optional, default is None).
W
Weilong Wu 已提交
2606
            For more information, please refer to :ref:`api_guide_Name`.
2607

W
Weilong Wu 已提交
2608
    Returns:
2609
        Tensor: The solution of a square system of linear equations with a unique solution for input 'x' and 'y'.
W
Weilong Wu 已提交
2610
        Its data type should be the same as that of `x`.
2611

W
Weilong Wu 已提交
2612 2613
    Examples:
    .. code-block:: python
2614

W
Weilong Wu 已提交
2615 2616 2617
        # a square system of linear equations:
        # 2*X0 + X1 = 9
        # X0 + 2*X1 = 8
2618

W
Weilong Wu 已提交
2619 2620
        import paddle
        import numpy as np
2621

W
Weilong Wu 已提交
2622 2623 2624 2625 2626
        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)
2627

W
Weilong Wu 已提交
2628 2629 2630
        print(out)
        # [2., 3.])
    """
Z
zhiboniu 已提交
2631
    if paddle.in_dynamic_mode():
W
Weilong Wu 已提交
2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643
        return _C_ops.solve(x, y)

    inputs = {"X": [x], "Y": [y]}
    helper = LayerHelper("solve", **locals())
    check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'solve')
    check_variable_and_dtype(y, 'y', ['float32', 'float64'], 'solve')
    out = helper.create_variable_for_type_inference(dtype=x.dtype)

    helper.append_op(
        type="solve", inputs={"X": x,
                              "Y": y}, outputs={"Out": out})
    return out
2644 2645


2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694
def triangular_solve(x,
                     y,
                     upper=True,
                     transpose=False,
                     unitriangular=False,
                     name=None):
    r"""
    Computes the solution of a system of equations with a triangular coefficient matrix `x` and
    multiple right-hand sides `y` .

    Input `x` and `y` is 2D matrices or batches of 2D matrices. If the inputs are batches, the outputs
    is also batches.

    Args:
        x (Tensor): The input triangular coefficient matrix. Its shape should be `[*, M, M]`, where `*` is zero or
            more batch dimensions. Its data type should be float32 or float64.
        y (Tensor): Multiple right-hand sides of system of equations. Its shape should be `[*, M, K]`, where `*` is 
            zero or more batch dimensions. Its data type should be float32 or float64.
        upper (bool, optional): Whether to solve the upper-triangular system of equations (default) or the lower-triangular 
            system of equations. Default: True.
        transpose (bool, optional): whether `x` should be transposed before calculation. Default: False.
        unitriangular (bool, optional): whether `x` is unit triangular. If True, the diagonal elements of `x` are assumed 
            to be 1 and not referenced from `x` . Default: False.
        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 the system of equations. Its data type should be the same as that of `x`.

    Examples:
    .. code-block:: python

        # a square system of linear equations:
        # x1 +   x2  +   x3 = 0
        #      2*x2  +   x3 = -9
        #               -x3 = 5

        import paddle
        import numpy as np

        x = paddle.to_tensor([[1, 1, 1], 
                              [0, 2, 1],
                              [0, 0,-1]], dtype="float64")
        y = paddle.to_tensor([[0], [-9], [5]], dtype="float64")
        out = paddle.linalg.triangular_solve(x, y, upper=True)

        print(out)
        # [7, -2, -5]
    """
Z
zhiboniu 已提交
2695
    if paddle.in_dynamic_mode():
2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718
        return _C_ops.triangular_solve(x, y, 'upper', upper, 'transpose',
                                       transpose, 'unitriangular',
                                       unitriangular)

    inputs = {"X": [x], "Y": [y]}
    helper = LayerHelper("triangular_solve", **locals())
    check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'triangular_solve')
    check_variable_and_dtype(y, 'y', ['float32', 'float64'], 'triangular_solve')
    out = helper.create_variable_for_type_inference(dtype=x.dtype)

    helper.append_op(
        type='triangular_solve',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs={
            'upper': upper,
            'transpose': transpose,
            'unitriangular': unitriangular
        })
    return out


Z
zhiboniu 已提交
2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751
def cholesky_solve(x, y, upper=False, name=None):
    r"""
    Solves a linear system of equations A @ X = B, given A's Cholesky factor matrix u and  matrix B.

    Input `x` and `y` is 2D matrices or batches of 2D matrices. If the inputs are batches, the outputs
    is also batches.

    Args:
        x (Tensor): The input matrix which is upper or lower triangular Cholesky factor of square matrix A. Its shape should be `[*, M, M]`, where `*` is zero or
            more batch dimensions. Its data type should be float32 or float64.
        y (Tensor): Multiple right-hand sides of system of equations. Its shape should be `[*, M, K]`, where `*` is 
            zero or more batch dimensions. Its data type should be float32 or float64.
        upper (bool, optional): whether to consider the Cholesky factor as a lower or upper triangular matrix. Default: False.
        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 the system of equations. Its data type is the same as that of `x`.

    Examples:
    .. code-block:: python

        import paddle

        u = paddle.to_tensor([[1, 1, 1], 
                                [0, 2, 1],
                                [0, 0,-1]], dtype="float64")
        b = paddle.to_tensor([[0], [-9], [5]], dtype="float64")
        out = paddle.linalg.cholesky_solve(b, u, upper=True)

        print(out)
        # [-2.5, -7, 9.5]
    """
Z
zhiboniu 已提交
2752
    if paddle.in_dynamic_mode():
Z
zhiboniu 已提交
2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768
        return _C_ops.cholesky_solve(x, y, 'upper', upper)

    helper = LayerHelper("cholesky_solve", **locals())
    check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'cholesky_solve')
    check_variable_and_dtype(y, 'y', ['float32', 'float64'], 'cholesky_solve')
    out = helper.create_variable_for_type_inference(dtype=x.dtype)

    helper.append_op(
        type='cholesky_solve',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs={'upper': upper})
    return out


2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795
def eigvalsh(x, UPLO='L', name=None):
    """
    Computes the eigenvalues of a 
    complex Hermitian (conjugate symmetric) or a real symmetric matrix.

    Args:
        x (Tensor): A tensor with shape :math:`[_, M, M]` , The data type of the input Tensor x
            should be one of float32, float64, complex64, complex128.
        UPLO(str, optional): Lower triangular part of a (‘L’, default) or the upper triangular part (‘U’).
        name(str, optional): The default value is None.  Normally there is no need for user to set this
            property.  For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor: The tensor eigenvalues in ascending order.

    Examples:
        .. code-block:: python

            import numpy as np
            import paddle

            x_data = np.array([[1, -2j], [2j, 5]])
            x = paddle.to_tensor(x_data)
            out_value = paddle.eigvalsh(x, UPLO='L')
            print(out_value)
            #[0.17157288, 5.82842712]
    """
Z
zhiboniu 已提交
2796
    if paddle.in_dynamic_mode():
2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810
        is_test = x.stop_gradient
        values, _ = _C_ops.eigvalsh(x, 'UPLO', UPLO, 'is_test', is_test)
        return values

    def __check_input(x, UPLO):
        x_shape = list(x.shape)
        if len(x.shape) < 2:
            raise ValueError(
                "Input(input) only support >=2 tensor, but received "
                "length of Input(input) is %s." % len(x.shape))
        if x_shape[-1] != x_shape[-2]:
            raise ValueError(
                "The input matrix must be batches of square matrices. But received x's dimention: {}".
                format(x_shape))
2811
        if UPLO != 'L' and UPLO != 'U':
2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833
            raise ValueError(
                "UPLO must be L or U. But received UPLO is: {}".format(UPLO))

    __check_input(x, UPLO)

    helper = LayerHelper('eigvalsh', **locals())
    check_variable_and_dtype(x, 'dtype',
                             ['float32', 'float64', 'complex64', 'complex128'],
                             'eigvalsh')

    out_value = helper.create_variable_for_type_inference(dtype=x.dtype)
    out_vector = helper.create_variable_for_type_inference(dtype=x.dtype)

    is_test = x.stop_gradient
    helper.append_op(
        type='eigvalsh',
        inputs={'X': x},
        outputs={'Eigenvalues': out_value,
                 'Eigenvectors': out_vector},
        attrs={'UPLO': UPLO,
               'is_test': is_test})
    return out_value
2834 2835


2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895
def lstsq(x, y, rcond=None, driver=None, name=None):
    """
    Computes a solution to
    the least squares problem of a system of linear equations.

    Args:
        x (Tensor): A tensor with shape ``(*, M, N)`` , the data type of the input Tensor ``x``
            should be one of float32, float64.
        y (Tensor): A tensor with shape ``(*, M, K)`` , the data type of the input Tensor ``y`` 
            should be one of float32, float64.
        rcond(float, optional): The default value is None. A float pointing number used to determine 
            the effective rank of ``x``. If ``rcond`` is None, it will be set to max(M, N) times the 
            machine precision of x_dtype.
        driver(str, optional): The default value is None. The name of LAPACK method to be used. For 
            CPU inputs the valid values are ‘gels’, ‘gelsy’, ‘gelsd, ‘gelss’. For CUDA input, the only 
            valid driver is ‘gels’. If ``driver`` is None, ‘gelsy’ is used for CPU inputs and ‘gels’ 
            for CUDA inputs.
        name(str, optional): The default value is None. Normally there is no need for user to set 
            this property. For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Tuple: A tuple of 4 Tensors which is (``solution``, ``residuals``, ``rank``, ``singular_values``). 
        ``solution`` is a tensor with shape ``(*, N, K)``, meaning the least squares solution. ``residuals`` 
        is a tensor with shape ``(*, K)``, meaning the squared residuals of the solutions, which is computed 
        when M > N and every matrix in ``x`` is full-rank, otherwise return an empty tensor. ``rank`` is a tensor 
        with shape ``(*)``, meaning the ranks of the matrices in ``x``, which is computed when ``driver`` in 
        (‘gelsy’, ‘gelsd’, ‘gelss’), otherwise return an empty tensor. ``singular_values`` is a tensor with 
        shape ``(*, min(M, N))``, meaning singular values of the matrices in ``x``, which is computed when 
        ``driver`` in (‘gelsd’, ‘gelss’), otherwise return an empty tensor.

    Examples:
        .. code-block:: python

            import paddle

            paddle.set_device("cpu")
            x = paddle.to_tensor([[1, 3], [3, 2], [5, 6.]])
            y = paddle.to_tensor([[3, 4, 6], [5, 3, 4], [1, 2, 1.]])
            results = paddle.linalg.lstsq(x, y, driver="gelsd")
            print(results[0])
            # [[ 0.78350395, -0.22165027, -0.62371236],
            # [-0.11340097,  0.78866047,  1.14948535]]
            print(results[1])
            # [19.81443405, 10.43814468, 30.56185532])
            print(results[2])
            # 2
            print(results[3])
            # [9.03455734, 1.54167950]

            x = paddle.to_tensor([[10, 2, 3], [3, 10, 5], [5, 6, 12.]])
            y = paddle.to_tensor([[4, 2, 9], [2, 0, 3], [2, 5, 3.]])
            results = paddle.linalg.lstsq(x, y, driver="gels")
            print(results[0])
            # [[ 0.39386186,  0.10230173,  0.93606132],
            # [ 0.10741687, -0.29028133,  0.11892585],
            # [-0.05115091,  0.51918161, -0.19948854]]
            print(results[1])
            # []
    """
    device = paddle.get_device()
2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910
    if device == "cpu":
        if driver not in (None, "gels", "gelss", "gelsd", "gelsy"):
            raise ValueError(
                "Only support valid driver is 'gels', 'gelss', 'gelsd', 'gelsy' or None for CPU inputs. But got {}".
                format(driver))
        driver = "gelsy" if driver is None else driver
    elif "gpu" in device:
        if driver not in (None, "gels"):
            raise ValueError(
                "Only support valid driver is 'gels' or None for CUDA inputs. But got {}".
                format(driver))
        driver = "gels" if driver is None else driver
    else:
        raise RuntimeError("Only support lstsq api for CPU or CUDA device.")

2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923
    if x.dtype == y.dtype and x.dtype in (paddle.float32, paddle.float64):
        pass
    else:
        raise ValueError(
            "Only support x and y have the same dtype such as 'float32' and 'float64'."
        )

    if rcond is None:
        if x.dtype == paddle.float32:
            rcond = 1e-7 * max(x.shape[-2], x.shape[-1])
        elif x.dtype == paddle.float64:
            rcond = 1e-15 * max(x.shape[-2], x.shape[-1])

Z
zhiboniu 已提交
2924
    if paddle.in_dynamic_mode():
2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008
        solution, rank, singular_values = _C_ops.lstsq(x, y, "rcond", rcond,
                                                       "driver", driver)
        if x.shape[-2] > x.shape[-1]:
            matmul_out = _varbase_creator(dtype=x.dtype)
            _C_ops.matmul(x, solution, matmul_out, 'trans_x', False, 'trans_y',
                          False)
            minus_out = _C_ops.elementwise_sub(matmul_out, y)
            pow_out = _C_ops.pow(minus_out, 'factor', 2)
            residuals = _C_ops.reduce_sum(pow_out, 'dim', [-2], 'keepdim',
                                          False, 'reduce_all', False)
        else:
            residuals = paddle.empty(shape=[0], dtype=x.dtype)

        if driver == "gels":
            rank = paddle.empty(shape=[0], dtype=paddle.int32)
            singular_values = paddle.empty(shape=[0], dtype=x.dtype)
        elif driver == "gelsy":
            singular_values = paddle.empty(shape=[0], dtype=x.dtype)

        return solution, residuals, rank, singular_values

    helper = LayerHelper('lstsq', **locals())
    check_variable_and_dtype(
        x, 'dtype', ['float32', 'float64', 'complex64', 'complex128'], 'lstsq')
    check_variable_and_dtype(
        y, 'dtype', ['float32', 'float64', 'complex64', 'complex128'], 'lstsq')

    solution = helper.create_variable_for_type_inference(dtype=x.dtype)
    residuals = helper.create_variable_for_type_inference(dtype=x.dtype)
    rank = helper.create_variable_for_type_inference(dtype=paddle.int32)
    singular_values = helper.create_variable_for_type_inference(dtype=x.dtype)

    helper.append_op(
        type='lstsq',
        inputs={'X': x,
                'Y': y},
        outputs={
            'Solution': solution,
            'Rank': rank,
            'SingularValues': singular_values
        },
        attrs={'rcond': rcond,
               'driver': driver})

    matmul_out = helper.create_variable_for_type_inference(dtype=x.dtype)
    minus_out = helper.create_variable_for_type_inference(dtype=x.dtype)
    pow_out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type='matmul_v2',
        inputs={'X': x,
                'Y': solution},
        outputs={'Out': matmul_out},
        attrs={
            'trans_x': False,
            'trans_y': False,
        })

    helper.append_op(
        type='elementwise_sub',
        inputs={'X': matmul_out,
                'Y': y},
        outputs={'Out': minus_out})

    helper.append_op(
        type='pow',
        inputs={'X': minus_out},
        outputs={'Out': pow_out},
        attrs={'factor': 2})

    helper.append_op(
        type='reduce_sum',
        inputs={'X': pow_out},
        outputs={'Out': residuals},
        attrs={'dim': [-2],
               'keep_dim': False,
               'reduce_all': False})

    if driver == "gels":
        rank = paddle.static.data(name='rank', shape=[0])
        singular_values = paddle.static.data(name='singular_values', shape=[0])
    elif driver == "gelsy":
        singular_values = paddle.static.data(name='singular_values', shape=[0])

    return solution, residuals, rank, singular_values