mindspore.nn.probability.distribution.Categorical

class mindspore.nn.probability.distribution.Categorical(probs=None, seed=None, dtype=mindspore.int32, name='Categorical')[source]

Create a categorical distribution parameterized by event probabilities.

Parameters
  • probs (Tensor, list, numpy.ndarray) – Event probabilities.

  • 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.

Supported Platforms:

Ascend GPU

Note

probs must have rank at least 1, values are proper probabilities and sum to 1.

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)
()
>>> # Examples of `sample`.
>>> # Args:
>>> #     shape (tuple): the shape of the sample. Default: ().
>>> #     probs (Tensor): event probabilities. Default: self.probs.
>>> ans = ca1.sample()
>>> print(ans.shape)
()
>>> ans = ca1.sample((2,3))
>>> print(ans.shape)
(2, 3)
>>> ans = ca1.sample((2,3), probs_b)
>>> print(ans.shape)
(2, 3)
>>> ans = ca2.sample((2,3), probs_a)
>>> print(ans.shape)
(2, 3)
property probs

Return the probability after casting to dtype.