PyTorch阶乘函数

Question

似乎没有计算阶乘的PyTorch函数。在PyTorch中有这样做的方法吗？我希望手动计算火炬中的泊松分布(我知道这是存在的：https://pytorch.org/docs/stable/generated/torch.poisson.html)，这个公式需要分母中的阶乘。

泊松分布：分布

Answer 1

我想你可以找到torch.jit._builtins.math.factorial ，但是 pytorch以及numpy和scipy (矮胖与枕骨的阶乘)都使用python的内建math.factorial。

import math

import numpy as np
import scipy as sp
import torch


print(torch.jit._builtins.math.factorial is math.factorial)
print(np.math.factorial is math.factorial)
print(sp.math.factorial is math.factorial)

True
True
True

但是，相比之下，scipy除了“主流”math.factorial之外，还包含非常“特殊”的阶乘函数scipy.special.factorial。与来自math模块的函数不同，它对数组进行操作：

from scipy import special

print(special.factorial is math.factorial)

False

# the all known factorial functions
factorials = (
    math.factorial,
    torch.jit._builtins.math.factorial,
    np.math.factorial,
    sp.math.factorial,
    special.factorial,
)

# Let's run some tests
tnsr = torch.tensor(3)

for fn in factorials:
    try:
        out = fn(tnsr)
    except Exception as err:
        print(fn.__name__, fn.__module__, ':', err)
    else:
        print(fn.__name__, fn.__module__, ':', out)

factorial math : 6
factorial math : 6
factorial math : 6
factorial math : 6
factorial scipy.special._basic : tensor(6., dtype=torch.float64)

tnsr = torch.tensor([1, 2, 3])

for fn in factorials:
    try:
        out = fn(tnsr)
    except Exception as err:
        print(fn.__name__, fn.__module__, ':', err)
    else:
        print(fn.__name__, fn.__module__, ':', out)

factorial math : only integer tensors of a single element can be converted to an index
factorial math : only integer tensors of a single element can be converted to an index
factorial math : only integer tensors of a single element can be converted to an index
factorial math : only integer tensors of a single element can be converted to an index
factorial scipy.special._basic : tensor([1., 2., 6.], dtype=torch.float64)

Answer 2

内置的math模块(文档)提供了一个函数，它将给定积分的阶乘作为int返回。

import math

x = math.factorial(5)
print(x)
print(type(x))

输出

120
<class 'int'>