[文档] torch.distributions.Categorical

这段时间看见分布式、并行之类的词语就害怕，结果这个是 distributions，分布，就是一些表征分布的函数们

签名：

torch.distributions.categorical.Categorical(probs=None, 
                                            logits=None, 
                                            validate_args=None)

Creates a categorical distribution parameterized by either probs or logits (but not both).
创建一个离散的类别分布，参数由 probs 和 logits, 二者其一指定

If probs is 1-dimensional with length-K, each element is the relative probability of sampling the class at that index.

If probs is N-dimensional, the first N-1 dimensions are treated as a batch of relative probability vectors.


如果 probs 是个一维的长度为K的张量，则每个元素是索引对应的类别的相对概率，将通过该概率进行采样
如 probs 是N维张量，前N-1维只会被视为响应的 Batch, 这个不好翻译，直接看例子把

第一个例子（官方Demo）：

>>> m = Categorical(torch.tensor([ 0.25, 0.25, 0.25, 0.25 ]))
>>> m.sample()  # 等概率的返回 0, 1, 2, 3

第二个例子：

probs = torch.FloatTensor([[0.05,0.1,0.85],[0.05,0.05,0.9]])
 
dist = Categorical(probs)
print(dist)
# Categorical(probs: torch.Size([2, 3]))
 
index = dist.sample()
print(index.numpy())
# 很大概率会是 [2 2]

如果 probs 变量的 -1 维度上求和不为1，Categorical 内部也会帮你归一化，然后再 sample

第二个例子：

>>> probs = torch.FloatTensor([[[0.05,0.1,0.85],[0.05,0.05,0.9]]])
>>> probs.shape
torch.Size([1, 2, 3])

>>> dist = Categorical(probs)
>>> index = dist.sample()
>>> index.shape
torch.Size([1, 2])

上边这个例子，用来理解这句话：

If probs is N-dimensional, the first N-1 dimensions are treated as a batch of relative probability vectors.

probs.shape 为 [X, Y, N, s]，index.shape 为 [X, Y, N]，Categorical 只在最后一维上采样

---

最后补一句，Categorical对象，有.entropy() 方法，用来计算熵

>>> probs = torch.FloatTensor([[[0.05,0.1,0.85],[0.05,0.05,0.9]]])
>>> dist = Categorical(probs)
>>> dist.entropy()
tensor([[0.5182, 0.3944]])