Source code for mindspore.nn.probability.distribution.categorical

# Copyright 2020 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""Categorical Distribution"""
import numpy as np
from mindspore import context
from mindspore.ops import operations as P
from mindspore.ops import composite as C
from mindspore.ops.functional import stop_gradient
from mindspore.ops.operations import _inner_ops as inner
from mindspore._checkparam import Validator
import mindspore.ops as ops
import mindspore.nn as nn
from mindspore.common import dtype as mstype
from .distribution import Distribution
from ._utils.utils import check_prob, check_sum_equal_one, check_rank,\
    check_distribution_name, raise_not_implemented_util
from ._utils.custom_ops import exp_generic, log_generic, broadcast_to


[docs]class Categorical(Distribution): """ Categorical distribution. A Categorical Distribution is a discrete distribution with the range {1, 2, ..., k} and the probability mass function as :math:`P(X = i) = p_i, i = 1, ..., k`. Args: probs (Tensor, list, numpy.ndarray): Event probabilities. Default: None. seed (int): The global seed is used in sampling. Global seed is used if it is None. Default: None. dtype (mindspore.dtype): The type of the event samples. Default: mstype.int32. name (str): The name of the distribution. Default: Categorical. Note: `probs` must have rank at least 1, values are proper probabilities and sum to 1. Raises: ValueError: When the sum of all elements in `probs` is not 1. Supported Platforms: ``Ascend`` ``GPU`` Examples: >>> import mindspore >>> import mindspore.nn as nn >>> import mindspore.nn.probability.distribution as msd >>> from mindspore import Tensor >>> # To initialize a Categorical distribution of probs [0.5, 0.5] >>> ca1 = msd.Categorical(probs=[0.2, 0.8], dtype=mindspore.int32) >>> # A Categorical distribution can be initialized without arguments. >>> # In this case, `probs` must be passed in through arguments during function calls. >>> ca2 = msd.Categorical(dtype=mindspore.int32) >>> # Here are some tensors used below for testing >>> value = Tensor([1, 0], dtype=mindspore.int32) >>> probs_a = Tensor([0.5, 0.5], dtype=mindspore.float32) >>> probs_b = Tensor([0.35, 0.65], dtype=mindspore.float32) >>> # Private interfaces of probability functions corresponding to public interfaces, including >>> # `prob`, `log_prob`, `cdf`, `log_cdf`, `survival_function`, and `log_survival`, are the same as follows. >>> # Args: >>> # value (Tensor): the value to be evaluated. >>> # probs (Tensor): event probabilities. Default: self.probs. >>> # Examples of `prob`. >>> # Similar calls can be made to other probability functions >>> # by replacing `prob` by the name of the function. >>> ans = ca1.prob(value) >>> print(ans.shape) (2,) >>> # Evaluate `prob` with respect to distribution b. >>> ans = ca1.prob(value, probs_b) >>> print(ans.shape) (2,) >>> # `probs` must be passed in during function calls. >>> ans = ca2.prob(value, probs_a) >>> print(ans.shape) (2,) >>> # Functions `mean`, `sd`, `var`, and `entropy` have the same arguments. >>> # Args: >>> # probs (Tensor): event probabilities. Default: self.probs. >>> # Examples of `mean`. `sd`, `var`, and `entropy` are similar. >>> ans = ca1.mean() # return 0.8 >>> print(ans.shape) (1,) >>> ans = ca1.mean(probs_b) >>> print(ans.shape) (1,) >>> # `probs` must be passed in during function calls. >>> ans = ca2.mean(probs_a) >>> print(ans.shape) (1,) >>> # Interfaces of `kl_loss` and `cross_entropy` are the same as follows: >>> # Args: >>> # dist (str): the name of the distribution. Only 'Categorical' is supported. >>> # probs_b (Tensor): event probabilities of distribution b. >>> # probs (Tensor): event probabilities of distribution a. Default: self.probs. >>> # Examples of `kl_loss`, `cross_entropy` is similar. >>> ans = ca1.kl_loss('Categorical', probs_b) >>> print(ans.shape) () >>> ans = ca1.kl_loss('Categorical', probs_b, probs_a) >>> print(ans.shape) () >>> # An additional `probs` must be passed in. >>> ans = ca2.kl_loss('Categorical', probs_b, probs_a) >>> print(ans.shape) () """ def __init__(self, probs=None, seed=None, dtype=mstype.int32, name="Categorical"): param = dict(locals()) param['param_dict'] = {'probs': probs} valid_dtype = mstype.uint_type + mstype.int_type + mstype.float_type Validator.check_type_name( "dtype", dtype, valid_dtype, type(self).__name__) super(Categorical, self).__init__(seed, dtype, name, param) self._probs = self._add_parameter(probs, 'probs') if self.probs is not None: check_rank(self.probs) check_prob(self.probs) check_sum_equal_one(probs) # update is_scalar_batch and broadcast_shape # drop one dimension if self.probs.shape[:-1] == (): self._is_scalar_batch = True self._broadcast_shape = self._broadcast_shape[:-1] self.argmax = P.ArgMaxWithValue(axis=-1) self.broadcast = broadcast_to self.cast = P.Cast() self.clip_by_value = ops.clip_by_value self.concat = P.Concat(-1) self.cumsum = P.CumSum() self.dtypeop = P.DType() self.exp = exp_generic self.expand_dim = P.ExpandDims() self.fill = P.Fill() self.gather = P.GatherNd() self.greater = P.Greater() self.issubclass = inner.IsSubClass() self.less = P.Less() # when the graph kernel mode is enable # use Log directly as akg will handle the corner cases self.log = P.Log() if context.get_context("enable_graph_kernel") else log_generic self.log_softmax = P.LogSoftmax() self.logicor = P.LogicalOr() self.logicand = P.LogicalAnd() self.multinomial = P.Multinomial(seed=self.seed) self.reshape = P.Reshape() self.reduce_sum = P.ReduceSum(keep_dims=True) self.select = P.Select() self.shape = P.Shape() self.softmax = P.Softmax() self.squeeze = P.Squeeze() self.squeeze_first_axis = P.Squeeze(0) self.squeeze_last_axis = P.Squeeze(-1) self.square = P.Square() self.transpose = P.Transpose() self.index_type = mstype.int32 self.nan = np.nan @property def probs(self): """ Return the probability after casting to dtype. Output: Tensor, the probs of the distribution. """ return self._probs def extend_repr(self): """Display instance object as string.""" if self.is_scalar_batch: s = 'probs = {}'.format(self.probs) else: s = 'batch_shape = {}'.format(self._broadcast_shape) return s def _get_dist_type(self): return "Categorical" def _get_dist_args(self, probs=None): if probs is not None: self.checktensor(probs, 'probs') else: probs = self.probs return (probs,) def _mean(self, probs=None): r""" .. math:: E[X] = \sum_{i=0}^{num_classes-1} i*p_i """ probs = self._check_param_type(probs) num_classes = self.shape(probs)[-1] index = nn.Range(0., num_classes, 1.)() return self.reduce_sum(index * probs, -1) def _mode(self, probs=None): probs = self._check_param_type(probs) index, _ = self.argmax(probs) mode = self.cast(index, self.dtype) return mode def _var(self, probs=None): r""" .. math:: VAR(X) = E[X^{2}] - (E[X])^{2} """ probs = self._check_param_type(probs) num_classes = self.shape(probs)[-1] index = nn.Range(0., num_classes, 1.)() return self.reduce_sum(self.square(index) * probs, -1) -\ self.square(self.reduce_sum(index * probs, -1)) def _entropy(self, probs=None): r""" Evaluate entropy. .. math:: H(X) = -\sum(logits * probs) """ probs = self._check_param_type(probs) logits = self.log(probs) return self.squeeze(-self.reduce_sum(logits * probs, -1)) def _kl_loss(self, dist, probs_b, probs=None): """ Evaluate KL divergence between Categorical distributions. Args: dist (str): The type of the distributions. Should be "Categorical" in this case. probs_b (Tensor): Event probabilities of distribution b. probs (Tensor): Event probabilities of distribution a. Default: self.probs. """ check_distribution_name(dist, 'Categorical') probs_b = self._check_value(probs_b, 'probs_b') probs_b = self.cast(probs_b, self.parameter_type) probs_a = self._check_param_type(probs) logits_a = self.log(probs_a) logits_b = self.log(probs_b) return self.squeeze(self.reduce_sum( self.softmax(logits_a) * (self.log_softmax(logits_a) - (self.log_softmax(logits_b))), -1)) def _cross_entropy(self, dist, probs_b, probs=None): """ Evaluate cross entropy between Categorical distributions. Args: dist (str): The type of the distributions. Should be "Categorical" in this case. probs_b (Tensor): Event probabilities of distribution b. probs (Tensor): Event probabilities of distribution a. Default: self.probs. """ check_distribution_name(dist, 'Categorical') return self._entropy(probs) + self._kl_loss(dist, probs_b, probs) def _log_prob(self, value, probs=None): r""" Evaluate log probability. Args: value (Tensor): The value to be evaluated. probs (Tensor): Event probabilities. Default: self.probs. """ value = self._check_value(value, 'value') probs = self._check_param_type(probs) logits = self.log(probs) # find the right integer to compute index # here we simulate casting to int but still keeping float dtype value = self.cast(value, self.dtypeop(probs)) zeros = self.fill(self.dtypeop(value), self.shape(value), 0.0) between_zero_neone = self.logicand(self.less(value, 0,), self.greater(value, -1.)) value = self.select(between_zero_neone, zeros, P.Floor()(value)) # handle the case when value is of shape () and probs is a scalar batch drop_dim = False if self.shape(value) == () and self.shape(probs)[:-1] == (): drop_dim = True # manually add one more dimension: () -> (1,) # drop this dimension before return value = self.expand_dim(value, -1) value = self.expand_dim(value, -1) broadcast_shape_tensor = logits * value broadcast_shape = self.shape(broadcast_shape_tensor) num_classes = broadcast_shape[-1] label_shape = broadcast_shape[:-1] # broadcasting logits and value # logit_pmf shape (num of labels, C) logits = self.broadcast(logits, broadcast_shape_tensor) value = self.broadcast(value, broadcast_shape_tensor)[..., :1] # flatten value to shape (number of labels, 1) # clip value to be in range from 0 to num_classes -1 and cast into int32 value = self.reshape(value, (-1, 1)) out_of_bound = self.squeeze_last_axis(self.logicor( self.less(value, 0.0), self.less(num_classes-1, value))) # deal with the case the there is only one class. value_clipped = self.clip_by_value(value, 0.0, num_classes - 1) value_clipped = self.cast(value_clipped, self.index_type) # create index from 0 ... NumOfLabels index = self.reshape(nn.Range(0, self.shape(value)[0], 1)(), (-1, 1)) index = self.concat((index, value_clipped)) # index into logit_pmf, fill in out_of_bound places with -inf # reshape into label shape N logits_pmf = self.gather(self.reshape( logits, (-1, num_classes)), index) nan = self.fill(self.dtypeop(logits_pmf), self.shape(logits_pmf), self.nan) logits_pmf = self.select(out_of_bound, nan, logits_pmf) ans = self.reshape(logits_pmf, label_shape) if drop_dim: return self.squeeze(ans) return ans def _cdf(self, value, probs=None): r""" Cumulative distribution function (cdf) of Categorical distributions. Args: value (Tensor): The value to be evaluated. probs (Tensor): Event probabilities. Default: self.probs. """ value = self._check_value(value, 'value') probs = self._check_param_type(probs) value = self.cast(value, self.dtypeop(probs)) zeros = self.fill(self.dtypeop(value), self.shape(value), 0.0) between_zero_neone = self.logicand( self.less(value, 0,), self.greater(value, -1.)) value = self.select(between_zero_neone, zeros, P.Floor()(value)) drop_dim = False if self.shape(value) == () and self.shape(probs)[:-1] == (): drop_dim = True value = self.expand_dim(value, -1) value = self.expand_dim(value, -1) broadcast_shape_tensor = probs * value broadcast_shape = self.shape(broadcast_shape_tensor) num_classes = broadcast_shape[-1] label_shape = broadcast_shape[:-1] probs = self.broadcast(probs, broadcast_shape_tensor) value = self.broadcast(value, broadcast_shape_tensor)[..., :1] # flatten value to shape (number of labels, 1) value = self.reshape(value, (-1, 1)) # drop one dimension to match cdf # clip value to be in range from 0 to num_classes -1 and cast into int32 less_than_zero = self.squeeze_last_axis(self.less(value, 0.0)) value_clipped = self.clip_by_value(value, 0.0, num_classes - 1) value_clipped = self.cast(value_clipped, self.index_type) index = self.reshape(nn.Range(0, self.shape(value)[0], 1)(), (-1, 1)) index = self.concat((index, value_clipped)) # reshape probs and fill less_than_zero places with 0 probs = self.reshape(probs, (-1, num_classes)) cdf = self.gather(self.cumsum(probs, 1), index) zeros = self.fill(self.dtypeop(cdf), self.shape(cdf), 0.0) cdf = self.select(less_than_zero, zeros, cdf) cdf = self.reshape(cdf, label_shape) if drop_dim: return self.squeeze(cdf) return cdf def _sample(self, shape=(), probs=None): """ Sampling. Args: shape (tuple): The shape of the sample. Default: (). probs (Tensor): Event probabilities. Default: self.probs. Returns: Tensor, shape is shape(probs)[:-1] + sample_shape """ if self.device_target == 'Ascend': raise_not_implemented_util('On d backend, sample', self.name) shape = self.checktuple(shape, 'shape') probs = self._check_param_type(probs) num_classes = self.shape(probs)[-1] batch_shape = self.shape(probs)[:-1] sample_shape = shape + batch_shape drop_dim = False if sample_shape == (): drop_dim = True sample_shape = (1,) probs_2d = self.reshape(probs, (-1, num_classes)) sample_tensor = self.fill(self.dtype, shape, 1.0) sample_tensor = self.reshape(sample_tensor, (-1, 1)) num_sample = self.shape(sample_tensor)[0] samples = C.multinomial(probs_2d, num_sample, seed=self.seed) samples = self.squeeze(self.transpose(samples, (1, 0))) samples = self.cast(self.reshape(samples, sample_shape), self.dtype) if drop_dim: return self.squeeze_first_axis(samples) samples = stop_gradient(samples) return samples