在 'reduceat' 中最快的 Python log-sum-exp

Question

作为统计编程包的一部分，我需要将对数转换值与 LogSumExp Function 一起添加。这比将未记录的值加在一起效率要低得多。

此外，我需要使用 numpy.ufunc.reduecat 功能将值相加。

我考虑过多种选择，代码如下：

（非对数比较space）使用numpy.add.reduceat
Numpy 的 ufunc 用于将记录的值加在一起：np.logaddexp.reduceat
具有以下 logsumexp 函数的手写 reduceat 函数：
- scipy's implemention of logsumexp
- Python 中的 logsumexp 函数（numba）
- Python 中的流式 logsumexp 函数（使用 numba）

def logsumexp_reduceat(arr, indices, logsum_exp_func):
    res = list()
    i_start = indices[0]
    for cur_index, i in enumerate(indices[1:]):
        res.append(logsum_exp_func(arr[i_start:i]))
        i_start = i

    res.append(logsum_exp_func(arr[i:]))
    return res

@numba.jit(nopython=True)
def logsumexp(X):
    r = 0.0
    for x in X:
        r += np.exp(x)  
    return np.log(r)

@numba.jit(nopython=True)
def logsumexp_stream(X):
    alpha = -np.Inf
    r = 0.0
    for x in X:
        if x != -np.Inf:
            if x <= alpha:
                r += np.exp(x - alpha)
            else:
                r *= np.exp(alpha - x)
                r += 1.0
                alpha = x
    return np.log(r) + alpha

arr = np.random.uniform(0,0.1, 10000)
log_arr = np.log(arr)
indices = sorted(np.random.randint(0, 10000, 100))

# approach 1
%timeit np.add.reduceat(arr, indices)
12.7 µs ± 503 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

# approach 2
%timeit np.logaddexp.reduceat(log_arr, indices)
462 µs ± 17.1 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

# approach 3, scipy function
%timeit logsum_exp_reduceat(arr, indices, scipy.special.logsumexp)
3.69 ms ± 273 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

# approach 3 handwritten logsumexp
%timeit logsumexp_reduceat(log_arr, indices, logsumexp)
139 µs ± 7.1 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

# approach 3 streaming logsumexp
%timeit logsumexp_reduceat(log_arr, indices, logsumexp_stream)
164 µs ± 10.9 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

timeit 结果表明，使用 numba 的手写 logsumexp 函数是最快的选项，但仍然比 numpy.add.reduceat.

慢 10 倍

几个问题：

是否有任何其他更快的方法（或对我提供的选项的调整）？例如，是否有使用查找的方法table 来计算 logsumexp 函数？
为什么 Sebastian Nowozin 的 "streaming logsumexp" 函数不比原始方法快？

Answer 1

还有一些改进空间

但永远不要指望 logsumexp 会像标准求和一样快，因为 exp 是一项相当昂贵的操作。

例子

import numpy as np

#from version 0.43 until 0.47 this has to be set before importing numba
#Bug: https://github.com/numba/numba/issues/4689
from llvmlite import binding
binding.set_option('SVML', '-vector-library=SVML')
import numba as nb

@nb.njit(fastmath=True,parallel=False)
def logsum_exp_reduceat(arr, indices):
    res = np.empty(indices.shape[0],dtype=arr.dtype)

    for i in nb.prange(indices.shape[0]-1):
        r = 0.
        for j in range(indices[i],indices[i+1]):
            r += np.exp(arr[j])  
        res[i]=np.log(r)

    r = 0.
    for j in range(indices[-1],arr.shape[0]):
        r += np.exp(arr[j])  
    res[-1]=np.log(r)
    return res

时间

#small example where parallelization doesn't make sense
arr = np.random.uniform(0,0.1, 10_000)
log_arr = np.log(arr)
#use arrays if possible
indices = np.sort(np.random.randint(0, 10_000, 100))

%timeit logsum_exp_reduceat(arr, indices)
#without parallelzation 22 µs ± 173 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
#with parallelization   84.7 µs ± 32.2 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

%timeit np.add.reduceat(arr, indices)
#4.46 µs ± 61.1 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

#large example where parallelization makes sense
arr = np.random.uniform(0,0.1, 1000_000)
log_arr = np.log(arr)
indices = np.sort(np.random.randint(0, 1000_000, 100))

%timeit logsum_exp_reduceat(arr, indices)
#without parallelzation 1.57 ms ± 14.1 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
#with parallelization   409 µs ± 14.7 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

%timeit np.add.reduceat(arr, indices)
#340 µs ± 11.1 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

在 'reduceat' 中最快的 Python log-sum-exp

Fastest Python log-sum-exp in a 'reduceat'

python

numpy

scipy

numba

numpy-ufunc