TensorFlow 概率采样需要很长时间

Question

我正在尝试使用 tfp 进行采样过程。从 beta 分布中抽取样本，并将结果作为概率输入从二项分布中抽取样本。它需要很长时间才能运行。

我应该以这种方式运行它还是有最佳方式？

'''

import tensorflow_probability as tfp
tfd = tfp.distributions

m = 100000 # sample size

### first sample from Beta distribution 
### and feed the result as probability to the binomial distribution sampling
s = tfd.Sample(
    tfd.Beta(2,2),
    sample_shape = m
)
phi = s.sample()

### Second sample from Binominal distribution 
### !!! it took forever to run...
s2 = tfd.Sample(
    tfd.Binomial(total_count=10, probs=phi),
    sample_shape = m
)

y = s2.sample() # not working well


### scipy code which works well:
from scipy import stats
m = 100000 # sample size
phi = stats.beta.rvs(2, 2, size = m)
y = stats.binom.rvs(10, phi, size = m)

'''

Answer 1

TFP 分布支持我们称为“批量形状”的概念。在这里，通过将probs=phi 与phi.shape = [100000] 一起提供，您实际上是在创建一个包含 100k 二项式的“批次”。然后你从那些样本中采样 100k 次，它试图创建 1e10 个样本，这需要一段时间！相反，试试这个：

m = 100000
s = tfd.Sample(
    tfd.Beta(2,2),
    sample_shape = m
)
phi = s.sample()

### Second sample from Binominal distribution 
s2 = tfd.Binomial(total_count=10, probs=phi)

y = s2.sample()

或者，使用tfd.BetaBinomial！

bb = tfd.BetaBinomial(total_count=10, concentration1=2, concentration0=2)
bb.sample(100000)

但最重要的是，看一下通过 TFP 的形状语义讨论的示例笔记本：https://www.tensorflow.org/probability/examples/Understanding_TensorFlow_Distributions_Shapes

干杯！