distribution.py 33.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
#   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.

# TODO: define the distribution functions 
# __all__ = ['Categorical',
#            'MultivariateNormalDiag',
#            'Normal',
#            'sampling_id',
#            'Uniform']
21 22 23 24 25 26 27

from __future__ import print_function

from .fluid.layers import control_flow
from .fluid.layers import tensor
from .fluid.layers import ops
from .fluid.layers import nn
28
from .fluid.layers import elementwise_mul, elementwise_div, elementwise_add, elementwise_sub
29
from .fluid import core
30
from .fluid.framework import in_dygraph_mode
P
pangyoki 已提交
31
from .tensor import arange, gather_nd, concat, multinomial
32 33 34 35 36 37
import math
import numpy as np
import warnings

from .fluid.data_feeder import convert_dtype, check_variable_and_dtype, check_type, check_dtype

P
pangyoki 已提交
38
__all__ = ['Distribution', 'Uniform', 'Normal', 'Categorical']
39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91


class Distribution(object):
    """
    The abstract base class for probability distributions. Functions are 
    implemented in specific distributions.
    """

    def __init__(self):
        super(Distribution, self).__init__()

    def sample(self):
        """Sampling from the distribution."""
        raise NotImplementedError

    def entropy(self):
        """The entropy of the distribution."""
        raise NotImplementedError

    def kl_divergence(self, other):
        """The KL-divergence between self distributions and other."""
        raise NotImplementedError

    def log_prob(self, value):
        """Log probability density/mass function."""
        raise NotImplementedError

    def probs(self, value):
        """Probability density/mass function."""
        raise NotImplementedError

    def _validate_args(self, *args):
        """
        Argument validation for distribution args
        Args:
            value (float, list, numpy.ndarray, Tensor)
        Raises
            ValueError: if one argument is Tensor, all arguments should be Tensor
        """
        is_variable = False
        is_number = False
        for arg in args:
            if isinstance(arg, tensor.Variable):
                is_variable = True
            else:
                is_number = True

        if is_variable and is_number:
            raise ValueError(
                'if one argument is Tensor, all arguments should be Tensor')

        return is_variable

92
    def _to_tensor(self, *args):
93 94 95 96 97 98 99 100 101 102 103 104 105 106
        """
        Argument convert args to Tensor

        Args:
            value (float, list, numpy.ndarray, Tensor)
        Returns:
            Tensor of args.
        """
        numpy_args = []
        variable_args = []
        tmp = 0.

        for arg in args:
            if isinstance(arg, float):
107 108 109 110 111 112
                arg = [arg]
            if not isinstance(arg, (list, np.ndarray, tensor.Variable)):
                raise TypeError(
                    "Type of input args must be float, list, numpy.ndarray or Tensor, but received type {}".
                    format(type(arg)))

113 114
            arg_np = np.array(arg)
            arg_dtype = arg_np.dtype
115 116 117 118 119 120 121
            if str(arg_dtype) != 'float32':
                if str(arg_dtype) != 'float64':
                    # "assign" op doesn't support float64. if dtype is float64, float32 variable will be generated
                    #  and converted to float64 later using "cast".
                    warnings.warn(
                        "data type of argument only support float32 and float64, your argument will be convert to float32."
                    )
122
                arg_np = arg_np.astype('float32')
123
            # tmp is used to support broadcast, it summarizes shapes of all the args and get the mixed shape.
124 125 126 127 128 129 130 131 132 133 134 135
            tmp = tmp + arg_np
            numpy_args.append(arg_np)

        dtype = tmp.dtype
        for arg in numpy_args:
            arg_broadcasted, _ = np.broadcast_arrays(arg, tmp)
            arg_variable = tensor.create_tensor(dtype=dtype)
            tensor.assign(arg_broadcasted, arg_variable)
            variable_args.append(arg_variable)

        return tuple(variable_args)

136 137 138 139 140 141
    def _check_values_dtype_in_probs(self, param, value):
        """
        Log_prob and probs methods have input ``value``, if value's dtype is different from param,
        convert value's dtype to be consistent with param's dtype.

        Args:
142
            param (Tensor): low and high in Uniform class, loc and scale in Normal class.
143 144 145 146 147 148 149 150 151 152 153 154 155
            value (Tensor): The input tensor.

        Returns:
            value (Tensor): Change value's dtype if value's dtype is different from param.
        """
        if in_dygraph_mode():
            if value.dtype != param.dtype and convert_dtype(
                    value.dtype) in ['float32', 'float64']:
                warnings.warn(
                    "dtype of input 'value' needs to be the same as parameters of distribution class. dtype of 'value' will be converted."
                )
                return core.ops.cast(value, 'in_dtype', value.dtype,
                                     'out_dtype', param.dtype)
156
            return value
157 158 159 160 161 162 163 164 165 166

        check_variable_and_dtype(value, 'value', ['float32', 'float64'],
                                 'log_prob')
        if value.dtype != param.dtype:
            warnings.warn(
                "dtype of input 'value' needs to be the same as parameters of distribution class. dtype of 'value' will be converted."
            )
            return tensor.cast(value, dtype=param.dtype)
        return value

167 168

class Uniform(Distribution):
169
    r"""Uniform distribution with `low` and `high` parameters.
170 171 172

    Mathematical Details

173
    The probability density function (pdf) is
174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192

    .. math::

        pdf(x; a, b) = \\frac{1}{Z}, \ a <=x <b

    .. math::

        Z = b - a

    In the above equation:

    * :math:`low = a`,
    * :math:`high = b`,
    * :math:`Z`: is the normalizing constant.

    The parameters `low` and `high` must be shaped in a way that supports
    [broadcasting](https://www.paddlepaddle.org.cn/documentation/docs/en/develop/beginners_guide/basic_concept/broadcasting_en.html) (e.g., `high - low` is a valid operation).

    Args:
193 194
        low(int|float|list|numpy.ndarray|Tensor): The lower boundary of uniform distribution.The data type is int, float, list, numpy.ndarray or Tensor
        high(int|float|list|numpy.ndarray|Tensor): The higher boundary of uniform distribution.The data type is int, float, list, numpy.ndarray or Tensor
195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243
        name(str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

    Examples:
        .. code-block:: python

          import numpy as np
          import paddle
          from paddle.distribution import Uniform

          paddle.disable_static()
          # Without broadcasting, a single uniform distribution [3, 4]:
          u1 = Uniform(low=3.0, high=4.0)
          # 2 distributions [1, 3], [2, 4]
          u2 = Uniform(low=[1.0, 2.0], high=[3.0, 4.0])
          # 4 distributions
          u3 = Uniform(low=[[1.0, 2.0], [3.0, 4.0]],
                    high=[[1.5, 2.5], [3.5, 4.5]])

          # With broadcasting:
          u4 = Uniform(low=3.0, high=[5.0, 6.0, 7.0])

          # Complete example
          value_npdata = np.array([0.8], dtype="float32")
          value_tensor = paddle.to_tensor(value_npdata)

          uniform = Uniform([0.], [2.])

          sample = uniform.sample([2])
          # a random tensor created by uniform distribution with shape: [2, 1]
          entropy = uniform.entropy()
          # [0.6931472] with shape: [1]
          lp = uniform.log_prob(value_tensor)
          # [-0.6931472] with shape: [1]
          p = uniform.probs(value_tensor)
          # [0.5] with shape: [1]
    """

    def __init__(self, low, high, name=None):
        if not in_dygraph_mode():
            check_type(low, 'low',
                       (int, float, np.ndarray, tensor.Variable, list),
                       'Uniform')
            check_type(high, 'high',
                       (int, float, np.ndarray, tensor.Variable, list),
                       'Uniform')

        self.all_arg_is_float = False
        self.batch_size_unknown = False
        self.name = name if name is not None else 'Uniform'
244
        self.dtype = 'float32'
245 246 247 248 249 250 251 252 253 254

        if isinstance(low, int):
            low = float(low)
        if isinstance(high, int):
            high = float(high)

        if self._validate_args(low, high):
            self.batch_size_unknown = True
            self.low = low
            self.high = high
255
            self.dtype = convert_dtype(low.dtype)
256 257 258
        else:
            if isinstance(low, float) and isinstance(high, float):
                self.all_arg_is_float = True
259 260 261 262 263 264 265 266
            if isinstance(
                    low,
                    np.ndarray) and str(low.dtype) in ['float32', 'float64']:
                self.dtype = low.dtype
            elif isinstance(
                    high,
                    np.ndarray) and str(high.dtype) in ['float32', 'float64']:
                self.dtype = high.dtype
267
            self.low, self.high = self._to_tensor(low, high)
268 269 270
            if self.dtype != convert_dtype(self.low.dtype):
                self.low = tensor.cast(self.low, dtype=self.dtype)
                self.high = tensor.cast(self.high, dtype=self.dtype)
271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291

    def sample(self, shape, seed=0):
        """Generate samples of the specified shape.

        Args:
          shape (list): 1D `int32`. Shape of the generated samples.
          seed (int): Python integer number.

        Returns:
          Tensor: A tensor with prepended dimensions shape.The data type is float32.

        """
        if not in_dygraph_mode():
            check_type(shape, 'shape', (list), 'sample')
            check_type(seed, 'seed', (int), 'sample')

        name = self.name + '_sample'
        batch_shape = list((self.low + self.high).shape)
        if self.batch_size_unknown:
            output_shape = shape + batch_shape
            zero_tmp = tensor.fill_constant_batch_size_like(
292
                self.low + self.high, batch_shape + shape, self.dtype, 0.)
293
            uniform_random_tmp = nn.uniform_random_batch_size_like(
294 295
                zero_tmp,
                zero_tmp.shape,
296
                dtype=self.dtype,
297 298 299 300 301 302 303 304 305 306
                min=0.,
                max=1.,
                seed=seed)
            zero_tmp_reshape = nn.reshape(zero_tmp, output_shape)
            uniform_random_tmp_reshape = nn.reshape(uniform_random_tmp,
                                                    output_shape)
            output = uniform_random_tmp_reshape * (
                zero_tmp_reshape + self.high - self.low)
            output = elementwise_add(output, self.low, name=name)
            return output
307 308 309
        else:
            output_shape = shape + batch_shape
            output = nn.uniform_random(
310 311
                output_shape, seed=seed, dtype=self.dtype) * (tensor.zeros(
                    output_shape, dtype=self.dtype) + (self.high - self.low))
312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328
            output = elementwise_add(output, self.low, name=name)
            if self.all_arg_is_float:
                return nn.reshape(output, shape, name=name)
            else:
                return output

    def log_prob(self, value):
        """Log probability density/mass function.

        Args:
          value (Tensor): The input tensor.

        Returns:
          Tensor: log probability.The data type is same with value.

        """
        name = self.name + '_log_prob'
329
        value = self._check_values_dtype_in_probs(self.low, value)
330
        if in_dygraph_mode():
331
            # ensure value in [low, high]
332 333
            lb_bool = self.low < value
            ub_bool = value < self.high
334 335

            lb = core.ops.cast(lb_bool, 'in_dtype', lb_bool.dtype, 'out_dtype',
336
                               value.dtype)
337
            ub = core.ops.cast(ub_bool, 'in_dtype', ub_bool.dtype, 'out_dtype',
338
                               value.dtype)
339
            return nn.log(lb * ub) - nn.log(self.high - self.low)
340

341 342
        lb_bool = self.low < value
        ub_bool = value < self.high
343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358
        lb = tensor.cast(lb_bool, dtype=value.dtype)
        ub = tensor.cast(ub_bool, dtype=value.dtype)
        return elementwise_sub(
            nn.log(lb * ub), nn.log(self.high - self.low), name=name)

    def probs(self, value):
        """Probability density/mass function.

        Args:
          value (Tensor): The input tensor.

        Returns:
          Tensor: probability.The data type is same with value.

        """
        name = self.name + '_probs'
359
        value = self._check_values_dtype_in_probs(self.low, value)
360 361 362
        if in_dygraph_mode():
            lb_bool = self.low < value
            ub_bool = value < self.high
363 364

            lb = core.ops.cast(lb_bool, 'in_dtype', lb_bool.dtype, 'out_dtype',
365
                               value.dtype)
366
            ub = core.ops.cast(ub_bool, 'in_dtype', ub_bool.dtype, 'out_dtype',
367
                               value.dtype)
368
            return (lb * ub) / (self.high - self.low)
369

370 371
        lb_bool = self.low < value
        ub_bool = value < self.high
372 373 374 375 376
        lb = tensor.cast(lb_bool, dtype=value.dtype)
        ub = tensor.cast(ub_bool, dtype=value.dtype)
        return elementwise_div((lb * ub), (self.high - self.low), name=name)

    def entropy(self):
377
        r"""Shannon entropy in nats.
378

379 380 381 382 383 384
        The entropy is

        .. math::

            entropy(low, high) = \\log (high - low)

385 386 387 388 389 390 391 392 393
        Returns:
          Tensor: Shannon entropy of uniform distribution.The data type is float32.

        """
        name = self.name + '_entropy'
        return nn.log(self.high - self.low, name=name)


class Normal(Distribution):
394
    r"""The Normal distribution with location `loc` and `scale` parameters.
395 396 397

    Mathematical details

398
    The probability density function (pdf) is
399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414

    .. math::

        pdf(x; \mu, \sigma) = \\frac{1}{Z}e^{\\frac {-0.5 (x - \mu)^2}  {\sigma^2} }

    .. math::

        Z = (2 \pi \sigma^2)^{0.5}

    In the above equation:

    * :math:`loc = \mu`: is the mean.
    * :math:`scale = \sigma`: is the std.
    * :math:`Z`: is the normalization constant.

    Args:
415 416
        loc(int|float|list|numpy.ndarray|Tensor): The mean of normal distribution.The data type is int, float, list, numpy.ndarray or Tensor.
        scale(int|float|list|numpy.ndarray|Tensor): The std of normal distribution.The data type is int, float, list, numpy.ndarray or Tensor.
417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468
        name(str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

    Examples:
        .. code-block:: python
          
          import numpy as np
          import paddle
          from paddle.distribution import Normal

          paddle.disable_static()
          # Define a single scalar Normal distribution.
          dist = Normal(loc=0., scale=3.)
          # Define a batch of two scalar valued Normals.
          # The first has mean 1 and standard deviation 11, the second 2 and 22.
          dist = Normal(loc=[1., 2.], scale=[11., 22.])
          # Get 3 samples, returning a 3 x 2 tensor.
          dist.sample([3])

          # Define a batch of two scalar valued Normals.
          # Both have mean 1, but different standard deviations.
          dist = Normal(loc=1., scale=[11., 22.])

          # Complete example
          value_npdata = np.array([0.8], dtype="float32")
          value_tensor = paddle.to_tensor(value_npdata)

          normal_a = Normal([0.], [1.])
          normal_b = Normal([0.5], [2.])
          sample = normal_a.sample([2])
          # a random tensor created by normal distribution with shape: [2, 1]
          entropy = normal_a.entropy()
          # [1.4189385] with shape: [1]
          lp = normal_a.log_prob(value_tensor)
          # [-1.2389386] with shape: [1]
          p = normal_a.probs(value_tensor)
          # [0.28969154] with shape: [1]
          kl = normal_a.kl_divergence(normal_b)
          # [0.34939718] with shape: [1]
    """

    def __init__(self, loc, scale, name=None):
        if not in_dygraph_mode():
            check_type(loc, 'loc',
                       (int, float, np.ndarray, tensor.Variable, list),
                       'Normal')
            check_type(scale, 'scale',
                       (int, float, np.ndarray, tensor.Variable, list),
                       'Normal')

        self.batch_size_unknown = False
        self.all_arg_is_float = False
        self.name = name if name is not None else 'Normal'
469
        self.dtype = 'float32'
470 471 472 473 474 475 476 477 478 479

        if isinstance(loc, int):
            loc = float(loc)
        if isinstance(scale, int):
            scale = float(scale)

        if self._validate_args(loc, scale):
            self.batch_size_unknown = True
            self.loc = loc
            self.scale = scale
480
            self.dtype = convert_dtype(loc.dtype)
481 482 483
        else:
            if isinstance(loc, float) and isinstance(scale, float):
                self.all_arg_is_float = True
484 485 486 487 488 489 490 491
            if isinstance(
                    loc,
                    np.ndarray) and str(loc.dtype) in ['float32', 'float64']:
                self.dtype = loc.dtype
            elif isinstance(
                    scale,
                    np.ndarray) and str(scale.dtype) in ['float32', 'float64']:
                self.dtype = scale.dtype
492
            self.loc, self.scale = self._to_tensor(loc, scale)
493 494 495
            if self.dtype != convert_dtype(self.loc.dtype):
                self.loc = tensor.cast(self.loc, dtype=self.dtype)
                self.scale = tensor.cast(self.scale, dtype=self.dtype)
496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517

    def sample(self, shape, seed=0):
        """Generate samples of the specified shape.

        Args:
          shape (list): 1D `int32`. Shape of the generated samples.
          seed (int): Python integer number.

        Returns:
          Tensor: A tensor with prepended dimensions shape.The data type is float32.

        """
        if not in_dygraph_mode():
            check_type(shape, 'shape', (list), 'sample')
            check_type(seed, 'seed', (int), 'sample')

        batch_shape = list((self.loc + self.scale).shape)
        name = self.name + '_sample'

        if self.batch_size_unknown:
            output_shape = shape + batch_shape
            zero_tmp = tensor.fill_constant_batch_size_like(
518
                self.loc + self.scale, batch_shape + shape, self.dtype, 0.)
519 520
            zero_tmp_reshape = nn.reshape(zero_tmp, output_shape)
            zero_tmp_shape = nn.shape(zero_tmp_reshape)
521
            normal_random_tmp = nn.gaussian_random(
522
                zero_tmp_shape, mean=0., std=1., seed=seed, dtype=self.dtype)
523 524 525
            output = normal_random_tmp * (zero_tmp_reshape + self.scale)
            output = elementwise_add(output, self.loc, name=name)
            return output
526 527
        else:
            output_shape = shape + batch_shape
528 529
            output = nn.gaussian_random(output_shape, mean=0., std=1., seed=seed, dtype=self.dtype) * \
                     (tensor.zeros(output_shape, dtype=self.dtype) + self.scale)
530 531 532 533 534 535 536
            output = elementwise_add(output, self.loc, name=name)
            if self.all_arg_is_float:
                return nn.reshape(output, shape, name=name)
            else:
                return output

    def entropy(self):
537
        r"""Shannon entropy in nats.
538

539 540 541 542 543 544 545 546 547 548
        The entropy is

        .. math::

            entropy(\sigma) = 0.5 \\log (2 \pi e \sigma^2)

        In the above equation:

        * :math:`scale = \sigma`: is the std.

549 550 551 552 553 554 555
        Returns:
          Tensor: Shannon entropy of normal distribution.The data type is float32.

        """
        name = self.name + '_entropy'
        batch_shape = list((self.loc + self.scale).shape)
        zero_tmp = tensor.fill_constant_batch_size_like(
556
            self.loc + self.scale, batch_shape, self.dtype, 0.)
557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572
        return elementwise_add(
            0.5 + zero_tmp,
            0.5 * math.log(2 * math.pi) + nn.log((self.scale + zero_tmp)),
            name=name)

    def log_prob(self, value):
        """Log probability density/mass function.

        Args:
          value (Tensor): The input tensor.

        Returns:
          Tensor: log probability.The data type is same with value.

        """
        name = self.name + '_log_prob'
573 574
        value = self._check_values_dtype_in_probs(self.loc, value)

575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592
        var = self.scale * self.scale
        log_scale = nn.log(self.scale)
        return elementwise_sub(
            -1. * ((value - self.loc) * (value - self.loc)) / (2. * var),
            log_scale + math.log(math.sqrt(2. * math.pi)),
            name=name)

    def probs(self, value):
        """Probability density/mass function.

        Args:
          value (Tensor): The input tensor.

        Returns:
          Tensor: probability.The data type is same with value.

        """
        name = self.name + '_probs'
593 594
        value = self._check_values_dtype_in_probs(self.loc, value)

595 596 597 598 599 600 601
        var = self.scale * self.scale
        return elementwise_div(
            ops.exp(-1. * ((value - self.loc) * (value - self.loc)) /
                    (2. * var)), (math.sqrt(2 * math.pi) * self.scale),
            name=name)

    def kl_divergence(self, other):
602
        r"""The KL-divergence between two normal distributions.
603

604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626
        The probability density function (pdf) is

        .. math::

            KL\_divergence(\mu_0, \sigma_0; \mu_1, \sigma_1) = 0.5 (ratio^2 + (\\frac{diff}{\sigma_1})^2 - 1 - 2 \\ln {ratio})

        .. math::

            ratio = \\frac{\sigma_0}{\sigma_1}
        
        .. math::

            diff = \mu_1 - \mu_0

        In the above equation:

        * :math:`loc = \mu_0`: is the mean of current Normal distribution.
        * :math:`scale = \sigma_0`: is the std of current Normal distribution.
        * :math:`loc = \mu_1`: is the mean of other Normal distribution.
        * :math:`scale = \sigma_1`: is the std of other Normal distribution.
        * :math:`ratio`: is the ratio of scales.
        * :math:`diff`: is the difference between means.

627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643
        Args:
            other (Normal): instance of Normal.

        Returns:
            Tensor: kl-divergence between two normal distributions.The data type is float32.

        """
        if not in_dygraph_mode():
            check_type(other, 'other', Normal, 'kl_divergence')

        name = self.name + '_kl_divergence'
        var_ratio = self.scale / other.scale
        var_ratio = (var_ratio * var_ratio)
        t1 = (self.loc - other.loc) / other.scale
        t1 = (t1 * t1)
        return elementwise_add(
            0.5 * var_ratio, 0.5 * (t1 - 1. - nn.log(var_ratio)), name=name)
P
pangyoki 已提交
644 645 646


class Categorical(Distribution):
647
    r"""
P
pangyoki 已提交
648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664
    Categorical distribution is a discrete probability distribution that 
    describes the possible results of a random variable that can take on 
    one of K possible categories, with the probability of each category 
    separately specified.

    The probability mass function (pmf) is:

    .. math::

        pmf(k; p_i) = \prod_{i=1}^{k} p_i^{[x=i]}

    In the above equation:

    * :math:`[x=i]` : it evaluates to 1 if :math:`x==i` , 0 otherwise.

    Args:
        logits(list|numpy.ndarray|Tensor): The logits input of categorical distribution. The data type is float32 or float64.
665
        name(str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.
P
pangyoki 已提交
666 667 668 669

    Examples:
        .. code-block:: python

670 671
            import paddle
            from paddle.distribution import Categorical
P
pangyoki 已提交
672

C
cnn 已提交
673
            paddle.seed(100) # on CPU device
674 675 676 677
            x = paddle.rand([6])
            print(x.numpy())
            # [0.5535528  0.20714243 0.01162981
            #  0.51577556 0.36369765 0.2609165 ]
P
pangyoki 已提交
678

C
cnn 已提交
679
            paddle.seed(200) # on CPU device
680 681 682 683
            y = paddle.rand([6])
            print(y.numpy())
            # [0.77663314 0.90824795 0.15685187
            #  0.04279523 0.34468332 0.7955718 ]
P
pangyoki 已提交
684

685 686
            cat = Categorical(x)
            cat2 = Categorical(y)
P
pangyoki 已提交
687

C
cnn 已提交
688
            paddle.seed(1000) # on CPU device
689 690 691
            cat.sample([2,3])
            # [[0, 0, 5],
            #  [3, 4, 5]]
P
pangyoki 已提交
692

693 694
            cat.entropy()
            # [1.77528]
P
pangyoki 已提交
695

696 697
            cat.kl_divergence(cat2)
            # [0.071952]
P
pangyoki 已提交
698

699 700 701 702 703 704
            value = paddle.to_tensor([2,1,3])
            cat.probs(value)
            # [0.00608027 0.108298 0.269656]

            cat.log_prob(value)
            # [-5.10271 -2.22287 -1.31061]
P
pangyoki 已提交
705 706 707 708 709 710

    """

    def __init__(self, logits, name=None):
        """
        Args:
711 712
            logits(list|numpy.ndarray|Tensor): The logits input of categorical distribution. The data type is float32 or float64.
            name(str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.
P
pangyoki 已提交
713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735
        """
        if not in_dygraph_mode():
            check_type(logits, 'logits', (np.ndarray, tensor.Variable, list),
                       'Categorical')

        self.name = name if name is not None else 'Categorical'
        self.dtype = 'float32'

        if self._validate_args(logits):
            self.logits = logits
            self.dtype = convert_dtype(logits.dtype)
        else:
            if isinstance(logits, np.ndarray) and str(
                    logits.dtype) in ['float32', 'float64']:
                self.dtype = logits.dtype
            self.logits = self._to_tensor(logits)[0]
            if self.dtype != convert_dtype(self.logits.dtype):
                self.logits = tensor.cast(self.logits, dtype=self.dtype)

    def sample(self, shape):
        """Generate samples of the specified shape.

        Args:
736
            shape (list): Shape of the generated samples.
P
pangyoki 已提交
737 738

        Returns:
739
            Tensor: A tensor with prepended dimensions shape.
P
pangyoki 已提交
740 741
        
        Examples:
742
            .. code-block:: python
P
pangyoki 已提交
743

744 745
                import paddle
                from paddle.distribution import Categorical
P
pangyoki 已提交
746

C
cnn 已提交
747
                paddle.seed(100) # on CPU device
748 749 750 751
                x = paddle.rand([6])
                print(x.numpy())
                # [0.5535528  0.20714243 0.01162981
                #  0.51577556 0.36369765 0.2609165 ]
P
pangyoki 已提交
752

753
                cat = Categorical(x)
P
pangyoki 已提交
754

C
cnn 已提交
755
                paddle.seed(1000) # on CPU device
756 757 758
                cat.sample([2,3])
                # [[0, 0, 5],
                #  [3, 4, 5]]
P
pangyoki 已提交
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

        """
        name = self.name + '_sample'
        if not in_dygraph_mode():
            check_type(shape, 'shape', (list), 'sample')

        num_samples = np.prod(np.array(shape))

        logits_shape = list(self.logits.shape)
        if len(logits_shape) > 1:
            sample_shape = shape + logits_shape[:-1]
            logits = nn.reshape(self.logits,
                                [np.prod(logits_shape[:-1]), logits_shape[-1]])
        else:
            sample_shape = shape
            logits = self.logits

        sample_index = multinomial(logits, num_samples, True)
        return nn.reshape(sample_index, sample_shape, name=name)

    def kl_divergence(self, other):
        """The KL-divergence between two Categorical distributions.

        Args:
            other (Categorical): instance of Categorical. The data type is float32.

        Returns:
786
            Tensor: kl-divergence between two Categorical distributions.
P
pangyoki 已提交
787 788
        
        Examples:
789
            .. code-block:: python
P
pangyoki 已提交
790

791 792 793
                import paddle
                from paddle.distribution import Categorical

C
cnn 已提交
794
                paddle.seed(100) # on CPU device
795 796 797 798
                x = paddle.rand([6])
                print(x.numpy())
                # [0.5535528  0.20714243 0.01162981
                #  0.51577556 0.36369765 0.2609165 ]
P
pangyoki 已提交
799

C
cnn 已提交
800
                paddle.seed(200) # on CPU device
801 802 803 804
                y = paddle.rand([6])
                print(y.numpy())
                # [0.77663314 0.90824795 0.15685187
                #  0.04279523 0.34468332 0.7955718 ]
P
pangyoki 已提交
805

806 807
                cat = Categorical(x)
                cat2 = Categorical(y)
P
pangyoki 已提交
808

809 810
                cat.kl_divergence(cat2)
                # [0.071952]
P
pangyoki 已提交
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

        """
        name = self.name + '_kl_divergence'
        if not in_dygraph_mode():
            check_type(other, 'other', Categorical, 'kl_divergence')

        logits = self.logits - nn.reduce_max(self.logits, dim=-1, keep_dim=True)
        other_logits = other.logits - nn.reduce_max(
            other.logits, dim=-1, keep_dim=True)
        e_logits = ops.exp(logits)
        other_e_logits = ops.exp(other_logits)
        z = nn.reduce_sum(e_logits, dim=-1, keep_dim=True)
        other_z = nn.reduce_sum(other_e_logits, dim=-1, keep_dim=True)
        prob = e_logits / z
        kl = nn.reduce_sum(
            prob * (logits - nn.log(z) - other_logits + nn.log(other_z)),
            dim=-1,
            keep_dim=True,
            name=name)

        return kl

    def entropy(self):
        """Shannon entropy in nats.

        Returns:
837
            Tensor: Shannon entropy of Categorical distribution. The data type is float32.
P
pangyoki 已提交
838 839
        
        Examples:
840
            .. code-block:: python
P
pangyoki 已提交
841

842 843
                import paddle
                from paddle.distribution import Categorical
P
pangyoki 已提交
844

C
cnn 已提交
845
                paddle.seed(100) # on CPU device
846 847 848 849
                x = paddle.rand([6])
                print(x.numpy())
                # [0.5535528  0.20714243 0.01162981
                #  0.51577556 0.36369765 0.2609165 ]
P
pangyoki 已提交
850

851
                cat = Categorical(x)
P
pangyoki 已提交
852

853 854
                cat.entropy()
                # [1.77528]
P
pangyoki 已提交
855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878

        """
        name = self.name + '_entropy'
        logits = self.logits - nn.reduce_max(self.logits, dim=-1, keep_dim=True)
        e_logits = ops.exp(logits)
        z = nn.reduce_sum(e_logits, dim=-1, keep_dim=True)
        prob = e_logits / z

        neg_entropy = nn.reduce_sum(
            prob * (logits - nn.log(z)), dim=-1, keep_dim=True)
        entropy = nn.scale(neg_entropy, scale=-1.0, name=name)
        return entropy

    def probs(self, value):
        """Probabilities of the given category (``value``).

        If ``logits`` is 2-D or higher dimension, the last dimension will be regarded as 
        category, and the others represents the different distributions.
        At the same time, if ``vlaue`` is 1-D Tensor, ``value`` will be broadcast to the 
        same number of distributions as ``logits``.
        If ``value`` is not 1-D Tensor, ``value`` should have the same number distributions
        with ``logits. That is, ``value[:-1] = logits[:-1]``.

        Args:
879
            value (Tensor): The input tensor represents the selected category index.
P
pangyoki 已提交
880 881

        Returns:
882
            Tensor: probability according to the category index.
P
pangyoki 已提交
883 884
        
        Examples:
885
            .. code-block:: python
P
pangyoki 已提交
886

887 888
                import paddle
                from paddle.distribution import Categorical
P
pangyoki 已提交
889

C
cnn 已提交
890
                paddle.seed(100) # on CPU device
891 892 893 894
                x = paddle.rand([6])
                print(x.numpy())
                # [0.5535528  0.20714243 0.01162981
                #  0.51577556 0.36369765 0.2609165 ]
P
pangyoki 已提交
895

896
                cat = Categorical(x)
P
pangyoki 已提交
897

898 899 900
                value = paddle.to_tensor([2,1,3])
                cat.probs(value)
                # [0.00608027 0.108298 0.269656]
P
pangyoki 已提交
901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944

        """
        name = self.name + '_probs'

        dist_sum = nn.reduce_sum(self.logits, dim=-1, keep_dim=True)
        prob = self.logits / dist_sum

        shape = list(prob.shape)
        value_shape = list(value.shape)
        if len(shape) == 1:
            num_value_in_one_dist = np.prod(value_shape)
            index_value = nn.reshape(value, [num_value_in_one_dist, 1])
            index = index_value
        else:
            num_dist = np.prod(shape[:-1])
            num_value_in_one_dist = value_shape[-1]
            prob = nn.reshape(prob, [num_dist, shape[-1]])
            if len(value_shape) == 1:
                value = nn.expand(value, [num_dist])
                value_shape = shape[:-1] + value_shape
            index_value = nn.reshape(value, [num_dist, -1, 1])
            if shape[:-1] != value_shape[:-1]:
                raise ValueError(
                    "shape of value {} must match shape of logits {}".format(
                        str(value_shape[:-1]), str(shape[:-1])))

            index_prefix = nn.unsqueeze(
                arange(
                    num_dist, dtype=index_value.dtype), axes=-1)
            index_prefix = nn.expand(index_prefix, [1, num_value_in_one_dist])
            index_prefix = nn.unsqueeze(index_prefix, axes=-1)

            if index_value.dtype != index_prefix.dtype:
                tensor.cast(index_prefix, dtype=index_value.dtype)
            index = concat([index_prefix, index_value], axis=-1)

        # value is the category index to search for the corresponding probability.
        select_prob = gather_nd(prob, index)
        return nn.reshape(select_prob, value_shape, name=name)

    def log_prob(self, value):
        """Log probabilities of the given category. Refer to ``probs`` method.

        Args:
945
            value (Tensor): The input tensor represents the selected category index.
P
pangyoki 已提交
946 947

        Returns:
948
            Tensor: Log probability.
P
pangyoki 已提交
949 950
        
        Examples:
951
            .. code-block:: python
P
pangyoki 已提交
952

953 954
                import paddle
                from paddle.distribution import Categorical
P
pangyoki 已提交
955

C
cnn 已提交
956
                paddle.seed(100) # on CPU device
957 958 959 960
                x = paddle.rand([6])
                print(x.numpy())
                # [0.5535528  0.20714243 0.01162981
                #  0.51577556 0.36369765 0.2609165 ]
P
pangyoki 已提交
961

962
                cat = Categorical(x)
P
pangyoki 已提交
963

964 965 966
                value = paddle.to_tensor([2,1,3])
                cat.log_prob(value)
                # [-5.10271 -2.22287 -1.31061]
P
pangyoki 已提交
967 968 969 970 971

        """
        name = self.name + '_log_prob'

        return nn.log(self.probs(value), name=name)