从 JAX 中的多元正态分布中抽样给出类型错误

Question

我正在尝试使用 JAX 从多元正态分布中生成样本：

import jax
import jax.numpy as jnp
import numpy as np

key = random.PRNGKey(0)
cov = np.array([[1.2, 0.4], [0.4, 1.0]])
mean = np.array([3,-1])
x1,x2 = jax.random.multivariate_normal(key, mean, cov, 5000).T

但是，当我运行代码时，出现以下错误：

TypeError                                 Traceback (most recent call last)
<ipython-input-25-1397bf923fa4> in <module>()
      2 cov = np.array([[1.2, 0.4], [0.4, 1.0]])
      3 mean = np.array([3,-1])
----> 4 x1,x2 = jax.random.multivariate_normal(key, mean, cov, 5000).T

1 frames
/usr/local/lib/python3.6/dist-packages/jax/core.py in canonicalize_shape(shape)
   1159          "got {}.")
   1160   if any(isinstance(x, Tracer) and isinstance(get_aval(x), ShapedArray)
-> 1161          and not isinstance(get_aval(x), ConcreteArray) for x in shape):
   1162     msg += ("\nIf using `jit`, try using `static_argnums` or applying `jit` to "
   1163             "smaller subfunctions.")

TypeError: 'int' object is not iterable

我不确定问题出在哪里，因为相同的语法适用于 Numpy 中的等效函数

Answer 1

在 jax.random 模块中，大多数形状必须明确为元组。因此，不要使用形状 5000，而是使用 (5000,):

x1,x2 = jax.random.multivariate_normal(key, mean, cov, (5000,)).T

从 JAX 中的多元正态分布中抽样给出类型错误

Sampling from multivariate normal distribution in JAX gives type error

python

numpy

normal-distribution

jax