linalg.py 79.5 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
H
huangxu96 已提交
17
from ..fluid.data_feeder import check_variable_and_dtype, check_type, check_dtype
18
from ..fluid.framework import in_dygraph_mode, _varbase_creator, Variable
19

20
from ..fluid.layers import transpose, cast  # noqa: F401
A
andyjpaddle 已提交
21 22
from ..fluid import layers
import paddle
23 24
from paddle.common_ops_import import core
from paddle.common_ops_import import VarDesc
W
wanghuancoder 已提交
25
from paddle import _C_ops
26

27 28
__all__ = []

29

S
ShenLiang 已提交
30
def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
31
    """
32 33
    Applies matrix multiplication to two tensors. `matmul` follows
    the complete broadcast rules,
S
ShenLiang 已提交
34
    and its behavior is consistent with `np.matmul`.
S
swtkiwi 已提交
35

S
ShenLiang 已提交
36 37
    Currently, the input tensors' number of dimensions can be any, `matmul` can be used to
    achieve the `dot`, `matmul` and `batchmatmul`.
38 39 40 41 42

    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
43 44
      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
45 46 47 48 49 50 51 52
      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.

53 54
    - 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
55
      After the matrix multiply, the prepended dimension is removed.
56 57

    - If the `x` is 2-dimensional and `y` is 1-dimensional,
S
ShenLiang 已提交
58 59
      the matrix-vector product is obtained.

60 61 62 63 64 65 66 67 68
    - 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
69
      out will be a (j, k, n, p) tensor.
70 71

    Args:
S
ShenLiang 已提交
72 73
        x (Tensor): The input tensor which is a Tensor.
        y (Tensor): The input tensor which is a Tensor.
74 75 76 77 78 79
        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
80
        Tensor: The output Tensor.
81 82 83

    Examples:

S
ShenLiang 已提交
84 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
    .. 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]
133 134

    """
S
ShenLiang 已提交
135 136
    op_type = 'matmul_v2'
    if in_dygraph_mode():
W
wanghuancoder 已提交
137
        op = getattr(_C_ops, op_type)
S
ShenLiang 已提交
138 139
        return op(x, y, 'trans_x', transpose_x, 'trans_y', transpose_y)

140
    attrs = {
S
ShenLiang 已提交
141 142
        'trans_x': transpose_x,
        'trans_y': transpose_y,
143 144 145 146 147
    }

    def __check_input(x, y):
        var_names = {'x': x, 'y': y}
        for name, val in var_names.items():
S
ShenLiang 已提交
148 149
            check_variable_and_dtype(
                val, name, ['float16', 'float32', 'float64'], 'matmul')
150 151 152

    __check_input(x, y)

S
ShenLiang 已提交
153
    helper = LayerHelper('matmul_v2', **locals())
154 155
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
S
ShenLiang 已提交
156
        type='matmul_v2',
157 158 159 160 161
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs=attrs)
    return out
Z
Zhang Ting 已提交
162 163


myq406450149's avatar
myq406450149 已提交
164
def norm(x, p='fro', axis=None, keepdim=False, name=None):
165
    """
S
swtkiwi 已提交
166

167 168 169
    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.

170 171 172 173 174 175
    .. 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.

176
    Args:
myq406450149's avatar
myq406450149 已提交
177
        x (Tensor): The input tensor could be N-D tensor, and the input data
178
            type could be float32 or float64.
myq406450149's avatar
myq406450149 已提交
179
        p (float|string, optional): Order of the norm. Supported values are `fro`, `0`, `1`, `2`,
180
            `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
181
            Default value is `fro`.
myq406450149's avatar
myq406450149 已提交
182 183
        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.
184
            If `axis < 0`, the dimension to norm operation is rank(input) + axis.
myq406450149's avatar
myq406450149 已提交
185
            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
186
            Defalut value is `None`.
187 188 189 190 191 192 193 194
        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
195
        Tensor: results of norm operation on the specified axis of input tensor,
196
        it's data type is the same as input's Tensor.
197

198 199
    Examples:
        .. code-block:: python
200

201
            import paddle
myq406450149's avatar
myq406450149 已提交
202 203 204 205 206 207 208 209
            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.]]]

210
            # compute frobenius norm along last two dimensions.
myq406450149's avatar
myq406450149 已提交
211 212 213
            out_fro = paddle.norm(x, p='fro', axis=[0,1])
            # out_fro.numpy() [17.435596 16.911535 16.7332   16.911535]

214 215
            # compute 2-order vector norm along last dimension.
            out_pnorm = paddle.norm(x, p=2, axis=-1)
myq406450149's avatar
myq406450149 已提交
216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233
            #out_pnorm.numpy(): [[21.118711  13.190906   5.477226]
            #                    [ 3.7416575 11.224972  19.131126]]

            # compute 2-order  norm along [0,1] dimension.
            out_pnorm = paddle.norm(x, p=2, axis=[0,1])
            #out_pnorm.numpy(): [17.435596 16.911535 16.7332   16.911535]

            # compute inf-order  norm
            out_pnorm = paddle.norm(x, p=np.inf)
            #out_pnorm.numpy()  = [12.]
            out_pnorm = paddle.norm(x, p=np.inf, axis=0)
            #out_pnorm.numpy(): [[12. 11. 10. 9.] [8. 7. 6. 7.] [8. 9. 10. 11.]]

            # compute -inf-order  norm
            out_pnorm = paddle.norm(x, p=-np.inf)
            #out_pnorm.numpy(): [0.]
            out_pnorm = paddle.norm(x, p=-np.inf, axis=0)
            #out_pnorm.numpy(): [[0. 1. 2. 3.] [4. 5. 6. 5.] [4. 3. 2. 1.]]
234 235
    """

myq406450149's avatar
myq406450149 已提交
236
    def frobenius_norm(input, dim=None, keepdim=False, name=None):
237 238 239 240 241 242 243 244 245 246 247
        """
        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!"
            )
myq406450149's avatar
myq406450149 已提交
248
        if in_dygraph_mode():
myq406450149's avatar
myq406450149 已提交
249
            if dim is None:
W
wanghuancoder 已提交
250 251 252 253
                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
254 255
        attrs = {'dim': dim, 'keep_dim': keepdim, 'reduce_all': False}
        if dim is None:
256 257 258 259 260
            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
261 262
        out = helper.create_variable_for_type_inference(
            dtype=helper.input_dtype())
263 264 265 266 267 268 269 270 271 272 273 274

        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
275
                    asvector=False,
276 277 278 279 280 281 282 283 284
                    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.
        """
myq406450149's avatar
myq406450149 已提交
285 286
        if in_dygraph_mode():
            if axis is None: axis = -1
W
wanghuancoder 已提交
287 288
            return _C_ops.p_norm(input, 'porder', porder, 'axis', axis,
                                 'keepdim', keepdim, 'asvector', asvector)
289 290 291 292
        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
293 294 295
        check_variable_and_dtype(input, 'input', ['float32', 'float64'],
                                 'p_norm')

296 297 298 299
        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
300
            'asvector': asvector,
301 302 303
            'epsilon': 1e-12,
        }
        helper = LayerHelper('p_norm', **locals())
myq406450149's avatar
myq406450149 已提交
304 305
        out = helper.create_variable_for_type_inference(
            dtype=helper.input_dtype())
306 307 308 309 310 311 312 313

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

myq406450149's avatar
myq406450149 已提交
314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342
    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):
343 344 345 346
        """
        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
347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380
        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

381 382 383
    if axis is None and p is not None:
        if isinstance(p, str):
            if p == "fro":
myq406450149's avatar
myq406450149 已提交
384
                return frobenius_norm(x, dim=axis, keepdim=keepdim, name=name)
385 386 387 388 389
            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
390 391 392 393 394 395
                x,
                porder=p,
                axis=axis,
                keepdim=keepdim,
                asvector=True,
                name=name)
396 397 398 399
        else:
            raise ValueError("only valid p type is string or float, found {}".
                             format(type(p)))

myq406450149's avatar
myq406450149 已提交
400 401
    if isinstance(axis, tuple):
        axis = list(axis)
402 403 404 405 406
    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
407 408 409 410 411 412 413 414 415 416 417 418 419 420
        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)):
421
            return vector_norm(
myq406450149's avatar
myq406450149 已提交
422 423 424 425 426 427
                x,
                axis=axis,
                porder=p,
                keepdim=keepdim,
                asvector=False,
                name=name)
428 429 430 431 432 433 434
        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
435 436 437
            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
438 439 440 441
        elif p == 0:
            raise ValueError(
                "just suport axis type int or list (length of list <=1) if p = 0, found {}".
                format(axis))
442
        else:
myq406450149's avatar
myq406450149 已提交
443 444
            return p_matrix_norm(
                x, porder=p, axis=axis, keepdim=keepdim, name=name)
445 446 447 448 449 450
    else:
        raise ValueError(
            "except axis type int or list (length of list <=2), found {}".
            format(axis))


Z
Zhang Ting 已提交
451
def dist(x, y, p=2):
452
    r"""
S
swtkiwi 已提交
453

Z
Zhang Ting 已提交
454
    This OP returns the p-norm of (x - y). It is not a norm in a strict sense, only as a measure
455 456
    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
457

458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480
    - 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
481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506

    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:
507 508
        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
509 510 511
        p (float, optional): The norm to be computed, its data type is float32 or float64. Default: 2.

    Returns:
512
        Tensor: Tensor that is the p-norm of (x - y).
Z
Zhang Ting 已提交
513 514 515 516 517 518 519

    Examples:
        .. code-block:: python

            import paddle
            import numpy as np

520 521 522 523
            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
524

525 526
            out = paddle.dist(x, y, 2)
            print(out) # out = [2.]
Z
Zhang Ting 已提交
527

528 529
            out = paddle.dist(x, y, float("inf"))
            print(out) # out = [2.]
Z
Zhang Ting 已提交
530

531 532
            out = paddle.dist(x, y, float("-inf"))
            print(out) # out = [0.]
Z
Zhang Ting 已提交
533 534 535 536 537 538 539 540 541 542 543 544 545
    """
    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
546 547


548 549 550 551 552 553
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:
554 555
        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``.
556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609
            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'))
610
            # a.numpy()
611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864
            # [[[ 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

        if in_dygraph_mode():
            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

        if in_dygraph_mode():
            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)

        if in_dygraph_mode():
            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):
        if in_dygraph_mode():
            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
865 866 867
def dot(x, y, name=None):
    """
    This operator calculates inner product for vectors.
868

L
liuwei1031 已提交
869
    .. note::
870 871
       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
872 873

    Parameters:
S
ShenLiang 已提交
874 875
        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
876 877
        name(str, optional): Name of the output. Default is None. It's used to print debug info for developers. Details: :ref:`api_guide_Name`

878
    Returns:
879
        Tensor: the calculated result Tensor.
880

L
liuwei1031 已提交
881 882 883 884 885 886
    Examples:

    .. code-block:: python

        import paddle
        import numpy as np
887 888 889

        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
890 891
        x = paddle.to_tensor(x_data)
        y = paddle.to_tensor(y_data)
892
        z = paddle.dot(x, y)
893
        print(z)
L
liuwei1031 已提交
894 895 896

    """
    op_type = 'dot'
897 898
    # skip var type check in dygraph mode to improve efficiency
    if in_dygraph_mode():
W
wanghuancoder 已提交
899
        op = getattr(_C_ops, op_type)
900 901
        return op(x, y)

L
liuwei1031 已提交
902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919
    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
920 921 922 923


def t(input, name=None):
    """
924 925
    Transpose <=2-D tensor.
    0-D and 1-D tensors are returned as it is and 2-D tensor is equal to
926
    the paddle.transpose function which perm dimensions set 0 and 1.
927

928
    Args:
929
        input (Tensor): The input Tensor. It is a N-D (N<=2) Tensor of data types float16, float32, float64, int32.
930
        name(str, optional): The default value is None.  Normally there is no need for
931 932
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`
    Returns:
933
        Tensor: A transposed n-D Tensor, with data type being float16, float32, float64, int32, int64.
934

935
    For Example:
936

937
        .. code-block:: text
938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953

             # 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])

954
     Examples:
955

956
        .. code-block:: python
957

958
            import paddle
959
            x = paddle.ones(shape=[2, 3], dtype='int32')
960
            x_transposed = paddle.t(x)
961 962
            print(x_transposed.shape)
            # [3, 2]
963 964 965 966 967 968 969 970 971 972 973
    """
    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))
    if in_dygraph_mode():
        if len(input.shape) == 1:
            return input
        # 2-D tensor
        perm = [1, 0]
W
wanghuancoder 已提交
974
        out, _ = _C_ops.transpose2(input, 'axis', perm)
975 976 977
        return out

    check_variable_and_dtype(
978 979
        input, 'input', ['float16', 'float32', 'float64', 'int32',
                         'int64'], 'transpose')
980 981 982 983 984 985 986 987 988 989 990 991 992 993

    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
994 995


996
def cross(x, y, axis=None, name=None):
997
    """
998
    Computes the cross product between two tensors along an axis.
999

1000 1001
    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.
1002

1003
    Args:
1004 1005
        x (Tensor): The first input tensor.
        y (Tensor): The second input tensor.
1006
        axis (int, optional): The axis along which to compute the cross product. It defaults to the first axis found with the length 3.
1007
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.
1008 1009

    Returns:
1010
        Tensor. A Tensor with same data type as `x`.
1011

1012 1013
    Examples:
        .. code-block:: python
1014

1015
            import paddle
1016

Z
Zhou Wei 已提交
1017 1018 1019 1020 1021 1022
            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]])
1023

1024 1025 1026 1027 1028 1029 1030 1031 1032
            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.]]
1033 1034
    """
    if in_dygraph_mode():
1035
        if axis is not None:
W
wanghuancoder 已提交
1036
            return _C_ops.cross(x, y, 'dim', axis)
1037
        else:
W
wanghuancoder 已提交
1038
            return _C_ops.cross(x, y)
1039

1040 1041
    helper = LayerHelper("cross", **locals())
    out = helper.create_variable_for_type_inference(x.dtype)
1042
    attrs = dict()
1043
    attrs['dim'] = axis
1044 1045 1046

    helper.append_op(
        type='cross',
1047 1048
        inputs={'X': x,
                'Y': y},
1049 1050 1051
        outputs={'Out': out},
        attrs=attrs)
    return out
1052 1053


1054
def cholesky(x, upper=False, name=None):
1055
    r"""
G
Guo Sheng 已提交
1056
    Computes the Cholesky decomposition of one symmetric positive-definite
1057 1058
    matrix or batches of symmetric positive-definite matrice.

G
Guo Sheng 已提交
1059 1060 1061 1062 1063 1064
    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:
1065
        x (Tensor): The input tensor. Its shape should be `[*, M, M]`,
G
Guo Sheng 已提交
1066 1067 1068 1069 1070 1071 1072
            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:
1073
        Tensor: A Tensor with same shape and data type as `x`. It represents \
G
Guo Sheng 已提交
1074
            triangular matrices generated by Cholesky decomposition.
1075

G
Guo Sheng 已提交
1076 1077 1078 1079 1080 1081
    Examples:
        .. code-block:: python

            import paddle
            import numpy as np

1082 1083 1084
            a = np.random.rand(3, 3)
            a_t = np.transpose(a, [1, 0])
            x_data = np.matmul(a, a_t) + 1e-03
1085
            x = paddle.to_tensor(x_data)
1086
            out = paddle.cholesky(x, upper=False)
1087
            print(out)
1088 1089 1090
            # [[1.190523   0.         0.        ]
            #  [0.9906703  0.27676893 0.        ]
            #  [1.25450498 0.05600871 0.06400121]]
G
Guo Sheng 已提交
1091 1092

    """
1093
    if in_dygraph_mode():
W
wanghuancoder 已提交
1094
        return _C_ops.cholesky(x, "upper", upper)
G
Guo Sheng 已提交
1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106
    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


1107 1108 1109 1110
def matrix_rank(x, tol=None, hermitian=False, name=None):
    r"""
    Computes the rank of a matrix.

1111
    The rank of a matrix is the number of singular values that are greater than the specified `tol` threshold when hermitian=False,
1112
    or the number of eigenvalues in absolute value that are greater than the specified `tol` threshold when hermitian=True.
1113 1114

    Args:
1115 1116 1117 1118
        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
1119
            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.
1120 1121
        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
1122
            the lower triangular of the matrix to compute.
1123 1124 1125 1126
        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.
1127

1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143
    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]]
1144

1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192
    """

    if in_dygraph_mode():
        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


1193 1194 1195 1196 1197 1198 1199 1200 1201
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
1202 1203
        x (Tensor): The input Tensor.
        y (Tensor): The input Tensor.
1204 1205 1206 1207
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.

    Returns:
Y
yaoxuefeng 已提交
1208
        Tensor: The product Tensor.
1209 1210

    Examples:
S
sunzhongkai588 已提交
1211 1212 1213
        .. code-block:: python

            import paddle
Y
yaoxuefeng 已提交
1214

S
sunzhongkai588 已提交
1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227
            # 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()
1228

1229
    """
Y
yaoxuefeng 已提交
1230 1231 1232 1233 1234 1235 1236 1237 1238 1239
    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))
1240 1241 1242 1243
    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))
1244

1245
    if in_dygraph_mode():
W
wanghuancoder 已提交
1246
        return _C_ops.bmm(x, y)
1247 1248

    helper = LayerHelper('bmm', **locals())
1249 1250 1251
    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
1252 1253 1254 1255


def histogram(input, bins=100, min=0, max=0):
    """
1256
    Computes the histogram of a tensor. The elements are sorted into equal width bins between min and max.
Q
Qi Li 已提交
1257 1258 1259
    If min and max are both zero, the minimum and maximum values of the data are used.

    Args:
1260
        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
1261 1262 1263 1264 1265 1266
            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:
1267
        Tensor: data type is int64, shape is (nbins,).
Q
Qi Li 已提交
1268

1269
    Examples:
Q
Qi Li 已提交
1270
        .. code-block:: python
1271

Q
Qi Li 已提交
1272
            import paddle
1273

1274
            inputs = paddle.to_tensor([1, 2, 1])
1275 1276
            result = paddle.histogram(inputs, bins=4, min=0, max=3)
            print(result) # [0, 2, 1, 0]
Q
Qi Li 已提交
1277 1278
    """
    if in_dygraph_mode():
W
wanghuancoder 已提交
1279
        return _C_ops.histogram(input, "bins", bins, "min", min, "max", max)
Q
Qi Li 已提交
1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292

    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
1293 1294 1295 1296 1297 1298 1299


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
1300
        x (Tensor): A tensor with shape :math:`[M, N]` , The data type of the input Tensor x
1301
            should be one of float32, float64.
F
furnace 已提交
1302
        vec (Tensor): A tensor with shape :math:`[N]` , The data type of the input Tensor x
1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325
            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)
    """
    if in_dygraph_mode():
W
wanghuancoder 已提交
1326
        out = _C_ops.mv(x, vec)
1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351
        return out

    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
1352 1353


H
huangxu96 已提交
1354 1355 1356 1357 1358 1359 1360 1361 1362
def det(x):
    """
    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.

1363
    Examples:
H
huangxu96 已提交
1364 1365 1366 1367 1368 1369 1370 1371 1372
        .. code-block:: python

        import paddle

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

        A = paddle.det(x)

        print(A)
1373

H
huangxu96 已提交
1374 1375
        # [ 0.02547996,  2.52317095, -6.15900707])

1376

H
huangxu96 已提交
1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405
    """
    if in_dygraph_mode():
        return core.ops.determinant(x)

    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


def slogdet(x):
    """
    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)
1406

H
huangxu96 已提交
1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417
    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.

1418
    Examples:
H
huangxu96 已提交
1419 1420 1421 1422 1423 1424 1425 1426 1427
    .. code-block:: python

        import paddle

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

        A = paddle.slogdet(x)

        print(A)
1428

H
huangxu96 已提交
1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456
        # [[ 1.        ,  1.        , -1.        ],
        # [-0.98610914, -0.43010661, -0.10872950]])

    """
    if in_dygraph_mode():
        return core.ops.slogdeterminant(x)

    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


1457 1458
def svd(x, full_matrices=False, name=None):
    r"""
1459 1460 1461 1462 1463
    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::
1464 1465
        X = U * diag(S) * VT

1466 1467
    Args:
        x (Tensor): The input tensor. Its shape should be `[..., N, M]`,
1468
            where `...` is zero or more batch dimensions. N and M can be arbitraty
1469 1470 1471 1472
            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,
1473
            which means shape of U is `[..., N, N]`, shape of V is `[..., M, M]`. K = min(M, N).
1474
            If full_matrices = False, svd op will use a economic method to store U and V.
1475
            which means shape of U is `[..., N, K]`, shape of V is `[..., M, K]`. K = min(M, N).
1476
        name (str, optional): Name for the operation (optional, default is None).
1477
            For more information, please refer to :ref:`api_guide_Name`.
1478 1479

    Returns:
1480
        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]`
1481

1482 1483 1484 1485
    Examples:
        .. code-block:: python

            import paddle
1486 1487 1488

            x = paddle.to_tensor([[1.0, 2.0], [1.0, 3.0], [4.0, 6.0]]).astype('float64')
            x = x.reshape([3, 2])
1489
            u, s, vh = paddle.linalg.svd(x)
1490 1491 1492 1493 1494
            print (u)
            #U = [[ 0.27364809, -0.21695147  ],
            #      [ 0.37892198, -0.87112408 ],
            #      [ 0.8840446 ,  0.44053933 ]]

1495
            print (s)
1496
            #S = [8.14753743, 0.78589688]
1497
            print (vh)
1498 1499
            #VT= [[ 0.51411221,  0.85772294],
            #     [ 0.85772294, -0.51411221]]
1500

1501
            # one can verify : U * S * VT == X
1502
            #                  U * UH == I
1503
            #                  V * VH == I
1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525
    """

    if in_dygraph_mode():
        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},
        attr=attrs, )
    return u, s, vh


1526 1527 1528
def matrix_power(x, n, name=None):
    r"""
    Computes the n-th power of a square matrix or a batch of square matrices.
1529

1530 1531 1532 1533 1534
    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}
1535

1536 1537 1538 1539
    Specifically,

    - If `n > 0`, it returns the matrix or a batch of matrices raised to the power
    of `n`.
1540

1541 1542 1543 1544 1545 1546 1547 1548 1549 1550
    - 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.
1551
        name (str, optional): Name for the operation (optional, default is None).
1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593
            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')
            print(paddle.matrix_power(x, 2))
            # [[6.  , 34. , 102.],
            #  [14. , 90. , 282.],
            #  [36. , 250., 804.]]

            print(paddle.matrix_power(x, 0))
            # [[1., 0., 0.],
            #  [0., 1., 0.],
            #  [0., 0., 1.]]

            print(paddle.matrix_power(x, -2))
            # [[ 12.91666667, -12.75000000,  2.83333333 ],
            #  [-7.66666667 ,  8.         , -1.83333333 ],
            #  [ 1.80555556 , -1.91666667 ,  0.44444444 ]]
    """
    if in_dygraph_mode():
        return core.ops.matrix_power(x, "n", n)

    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
1594 1595


1596 1597 1598
def eigvals(x, name=None):
    """
    Compute the eigenvalues of one or more general matrices.
1599 1600 1601

    Warning:
        The gradient kernel of this operator does not yet developed.
1602 1603 1604 1605
        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.
1606
            Its shape should be `[*, M, M]`, where `*` is zero or more batch dimensions.
1607
            Its data type should be float32, float64, complex64, or complex128.
1608
        name (str, optional): Name for the operation (optional, default is None).
1609 1610 1611
            For more information, please refer to :ref:`api_guide_Name`.

    Returns:
1612
        Tensor: A tensor containing the unsorted eigenvalues which has the same batch dimensions with `x`.
1613 1614 1615 1616 1617 1618
            The eigenvalues are complex-valued even when `x` is real.

    Examples:
        .. code-block:: python

            import paddle
1619

1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632
            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',
1633 1634
                             ['float32', 'float64', 'complex64',
                              'complex128'], 'eigvals')
1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655

    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))

    if in_dygraph_mode():
        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


1656 1657 1658 1659
def multi_dot(x, name=None):
    """
    Multi_dot is an operator that calculates multiple matrix multiplications.

1660
    Supports inputs of float16(only GPU support), float32 and float64 dtypes. This function does not
1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701
    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)
1702
        out = paddle.linalg.multi_dot([A, B])
1703 1704 1705 1706 1707 1708 1709 1710 1711 1712
        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)
1713
        out = paddle.linalg.multi_dot([A, B, C])
1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733
        print(out.numpy().shape)
        # [10, 7]

    """
    if in_dygraph_mode():
        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
1734 1735 1736 1737


def eigh(x, UPLO='L', name=None):
    """
1738
    Compute the eigenvalues and eigenvectors of a
1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761
    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)
1762
            out_value, out_vector = paddle.linalg.eigh(x, UPLO='L')
1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782
            print(out_value)
            #[0.17157288, 5.82842712]
            print(out_vector)
            #[(-0.9238795325112867+0j), (-0.3826834323650898+0j)],
            #[ 0.3826834323650898j    , -0.9238795325112867j    ]]

    """
    if in_dygraph_mode():
        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))
1783
        if UPLO != 'L' and UPLO != 'U':
1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802
            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
1803 1804 1805 1806


def pinv(x, rcond=1e-15, hermitian=False, name=None):
    r"""
1807
    Calculate pseudo inverse via SVD(singular value decomposition)
A
andyjpaddle 已提交
1808 1809 1810 1811 1812 1813 1814 1815 1816 1817
    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)
1818

A
andyjpaddle 已提交
1819 1820 1821
    If x is hermitian or symmetric matrix, svd will be replaced with eigh.

    Args:
1822 1823 1824
        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
1825 1826 1827 1828
            float32 or float64 or complex64 or complex128. When data
            type is complex64 or cpmplex128, hermitian should be set
            True.

1829 1830 1831 1832
        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
1833
            if complex or symmetric if real. Default: False.
1834 1835

        name(str|None): A name for this layer(optional). If set None,
A
andyjpaddle 已提交
1836
            the layer will be named automatically.
1837

A
andyjpaddle 已提交
1838
    Returns:
1839
        Tensor: The tensor with same data type with x. it represents
A
andyjpaddle 已提交
1840
        pseudo inverse of x. Its shape should be (*, n, m).
1841

A
andyjpaddle 已提交
1842 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 1920 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
    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 ;
    """

    if in_dygraph_mode():
        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(
2001 2002
                x, 'dtype', ['float32', 'float64', 'complex64',
                             'complex128'], 'pinv')
A
andyjpaddle 已提交
2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 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

            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
2075 2076 2077 2078 2079 2080 2081


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:
2082

W
Weilong Wu 已提交
2083 2084 2085 2086
    .. math::
        Out = X^-1 * Y
    Specifically,
    - This system of linear equations has one solution if and only if input 'X' is invertible.
2087

W
Weilong Wu 已提交
2088 2089 2090 2091 2092
    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.
2093
        name(str, optional): Name for the operation (optional, default is None).
W
Weilong Wu 已提交
2094
            For more information, please refer to :ref:`api_guide_Name`.
2095

W
Weilong Wu 已提交
2096
    Returns:
2097
        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
2098
        Its data type should be the same as that of `x`.
2099

W
Weilong Wu 已提交
2100 2101
    Examples:
    .. code-block:: python
2102

W
Weilong Wu 已提交
2103 2104 2105
        # a square system of linear equations:
        # 2*X0 + X1 = 9
        # X0 + 2*X1 = 8
2106

W
Weilong Wu 已提交
2107 2108
        import paddle
        import numpy as np
2109

W
Weilong Wu 已提交
2110 2111 2112 2113 2114
        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)
2115

W
Weilong Wu 已提交
2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131
        print(out)
        # [2., 3.])
    """
    if in_dygraph_mode():
        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