Random Distribution Generator NDArray API

Overview

This document lists the random distribution generator routines of the n-dimensional array package:

mxnet.ndarray.random Random distribution generator NDArray API of MXNet.

The Random Distribution Generator NDArray API, defined in the ndarray.random package, provides imperative random distribution generator operations on CPU/GPU.

In the rest of this document, we list routines provided by the ndarray.random package.

Random Distribution Generator

exponential Draw samples from an exponential distribution.
gamma Draw random samples from a gamma distribution.
generalized_negative_binomial Draw random samples from a generalized negative binomial distribution.
negative_binomial Draw random samples from a negative binomial distribution.
normal Draw random samples from a normal (Gaussian) distribution.
poisson Draw random samples from a Poisson distribution.
uniform Draw random samples from a uniform distribution.
multinomial Concurrent sampling from multiple multinomial distributions.
shuffle Shuffle the elements randomly.
mxnet.random.seed Seeds the random number generators in MXNet.

API Reference

Random distribution generator NDArray API of MXNet.

mxnet.ndarray.random.uniform(low=0, high=1, shape=_Null, dtype=_Null, ctx=None, out=None, **kwargs)[source]

Draw random samples from a uniform distribution.

Samples are uniformly distributed over the half-open interval [low, high) (includes low, but excludes high).

Parameters:
  • low (float or NDArray) – Lower boundary of the output interval. All values generated will be greater than or equal to low. The default value is 0.
  • high (float or NDArray) – Upper boundary of the output interval. All values generated will be less than high. The default value is 1.0.
  • shape (int or tuple of ints) – The number of samples to draw. If shape is, e.g., (m, n) and low and high are scalars, output shape will be (m, n). If low and high are NDArrays with shape, e.g., (x, y), then output will have shape (x, y, m, n), where m*n samples are drawn for each [low, high) pair.
  • dtype ({'float16','float32', 'float64'}) – Data type of output samples. Default is ‘float32’
  • ctx (Context) – Device context of output. Default is current context. Overridden by low.context when low is an NDArray.
  • out (NDArray) – Store output to an existing NDArray.

Examples

>>> mx.nd.random.uniform(0, 1)
[ 0.54881352]

>>> mx.nd.random.uniform(0, 1, ctx=mx.gpu(0))
[ 0.92514056]

>>> mx.nd.random.uniform(-1, 1, shape=(2,))
[ 0.71589124  0.08976638]

>>> low = mx.nd.array([1,2,3])
>>> high = mx.nd.array([2,3,4])
>>> mx.nd.random.uniform(low, high, shape=2)
[[ 1.78653979  1.93707538]
 [ 2.01311183  2.37081361]
 [ 3.30491424  3.69977832]]

mxnet.ndarray.random.normal(loc=0, scale=1, shape=_Null, dtype=_Null, ctx=None, out=None, **kwargs)[source]

Draw random samples from a normal (Gaussian) distribution.

Samples are distributed according to a normal distribution parametrized by loc (mean) and scale (standard deviation).

Parameters:
  • loc (float or NDArray) – Mean (centre) of the distribution.
  • scale (float or NDArray) – Standard deviation (spread or width) of the distribution.
  • shape (int or tuple of ints) – The number of samples to draw. If shape is, e.g., (m, n) and loc and scale are scalars, output shape will be (m, n). If loc and scale are NDArrays with shape, e.g., (x, y), then output will have shape (x, y, m, n), where m*n samples are drawn for each [loc, scale) pair.
  • dtype ({'float16','float32', 'float64'}) – Data type of output samples. Default is ‘float32’
  • ctx (Context) – Device context of output. Default is current context. Overridden by loc.context when loc is an NDArray.
  • out (NDArray) – Store output to an existing NDArray.

Examples

>>> mx.nd.random.normal(0, 1)
[ 2.21220636]

>>> mx.nd.random.normal(0, 1, ctx=mx.gpu(0))
[ 0.29253659]

>>> mx.nd.random.normal(-1, 1, shape=(2,))
[-0.2259962  -0.51619542]

>>> loc = mx.nd.array([1,2,3])
>>> scale = mx.nd.array([2,3,4])
>>> mx.nd.random.normal(loc, scale, shape=2)
[[ 0.55912292  3.19566321]
 [ 1.91728961  2.47706747]
 [ 2.79666662  5.44254589]]

mxnet.ndarray.random.poisson(lam=1, shape=_Null, dtype=_Null, ctx=None, out=None, **kwargs)[source]

Draw random samples from a Poisson distribution.

Samples are distributed according to a Poisson distribution parametrized by lambda (rate). Samples will always be returned as a floating point data type.

Parameters:
  • lam (float or NDArray) – Expectation of interval, should be >= 0.
  • shape (int or tuple of ints) – The number of samples to draw. If shape is, e.g., (m, n) and lam is a scalar, output shape will be (m, n). If lam is an NDArray with shape, e.g., (x, y), then output will have shape (x, y, m, n), where m*n samples are drawn for each entry in lam.
  • dtype ({'float16','float32', 'float64'}) – Data type of output samples. Default is ‘float32’
  • ctx (Context) – Device context of output. Default is current context. Overridden by lam.context when lam is an NDArray.
  • out (NDArray) – Store output to an existing NDArray.

Examples

>>> mx.nd.random.poisson(1)
[ 1.]

>>> mx.nd.random.poisson(1, shape=(2,))
[ 0.  2.]

>>> lam = mx.nd.array([1,2,3])
>>> mx.nd.random.poisson(lam, shape=2)
[[ 1.  3.]
 [ 3.  2.]
 [ 2.  3.]]

mxnet.ndarray.random.exponential(scale=1, shape=_Null, dtype=_Null, ctx=None, out=None, **kwargs)[source]

Draw samples from an exponential distribution.

Its probability density function is

\[f(x; \frac{1}{\beta}) = \frac{1}{\beta} \exp(-\frac{x}{\beta}),\]

for x > 0 and 0 elsewhere. beta is the scale parameter, which is the inverse of the rate parameter lambda = 1/beta.

Parameters:
  • scale (float or NDArray) – The scale parameter, beta = 1/lambda.
  • shape (int or tuple of ints) – The number of samples to draw. If shape is, e.g., (m, n) and scale is a scalar, output shape will be (m, n). If scale is an NDArray with shape, e.g., (x, y), then output will have shape (x, y, m, n), where m*n samples are drawn for each entry in scale.
  • dtype ({'float16','float32', 'float64'}) – Data type of output samples. Default is ‘float32’
  • ctx (Context) – Device context of output. Default is current context. Overridden by scale.context when scale is an NDArray.
  • out (NDArray) – Store output to an existing NDArray.

Examples

>>> mx.nd.random.exponential(1)
[ 0.79587454]

>>> mx.nd.random.exponential(1, shape=(2,))
[ 0.89856035  1.25593066]

>>> scale = mx.nd.array([1,2,3])
>>> mx.nd.random.exponential(scale, shape=2)
[[  0.41063145   0.42140478]
 [  2.59407091  10.12439728]
 [  2.42544937   1.14260709]]

mxnet.ndarray.random.gamma(alpha=1, beta=1, shape=_Null, dtype=_Null, ctx=None, out=None, **kwargs)[source]

Draw random samples from a gamma distribution.

Samples are distributed according to a gamma distribution parametrized by alpha (shape) and beta (scale).

Parameters:
  • alpha (float or NDArray) – The shape of the gamma distribution. Should be greater than zero.
  • beta (float or NDArray) – The scale of the gamma distribution. Should be greater than zero. Default is equal to 1.
  • shape (int or tuple of ints) – The number of samples to draw. If shape is, e.g., (m, n) and alpha and beta are scalars, output shape will be (m, n). If alpha and beta are NDArrays with shape, e.g., (x, y), then output will have shape (x, y, m, n), where m*n samples are drawn for each [alpha, beta) pair.
  • dtype ({'float16','float32', 'float64'}) – Data type of output samples. Default is ‘float32’
  • ctx (Context) – Device context of output. Default is current context. Overridden by alpha.context when alpha is an NDArray.
  • out (NDArray) – Store output to an existing NDArray.

Examples

>>> mx.nd.random.gamma(1, 1)
[ 1.93308783]

>>> mx.nd.random.gamma(1, 1, shape=(2,))
[ 0.48216391  2.09890771]

>>> alpha = mx.nd.array([1,2,3])
>>> beta = mx.nd.array([2,3,4])
>>> mx.nd.random.gamma(alpha, beta, shape=2)
[[  3.24343276   0.94137681]
 [  3.52734375   0.45568955]
 [ 14.26264095  14.0170126 ]]

mxnet.ndarray.random.negative_binomial(k=1, p=1, shape=_Null, dtype=_Null, ctx=None, out=None, **kwargs)[source]

Draw random samples from a negative binomial distribution.

Samples are distributed according to a negative binomial distribution parametrized by k (limit of unsuccessful experiments) and p (failure probability in each experiment). Samples will always be returned as a floating point data type.

Parameters:
  • k (float or NDArray) – Limit of unsuccessful experiments, > 0.
  • p (float or NDArray) – Failure probability in each experiment, >= 0 and <=1.
  • shape (int or tuple of ints) – The number of samples to draw. If shape is, e.g., (m, n) and k and p are scalars, output shape will be (m, n). If k and p are NDArrays with shape, e.g., (x, y), then output will have shape (x, y, m, n), where m*n samples are drawn for each [k, p) pair.
  • dtype ({'float16','float32', 'float64'}) – Data type of output samples. Default is ‘float32’
  • ctx (Context) – Device context of output. Default is current context. Overridden by k.context when k is an NDArray.
  • out (NDArray) – Store output to an existing NDArray.

Examples

>>> mx.nd.random.negative_binomial(10, 0.5)
[ 4.]

>>> mx.nd.random.negative_binomial(10, 0.5, shape=(2,))
[ 3.  4.]

>>> k = mx.nd.array([1,2,3])
>>> p = mx.nd.array([0.2,0.4,0.6])
>>> mx.nd.random.negative_binomial(k, p, shape=2)
[[ 3.  2.]
 [ 4.  4.]
 [ 0.  5.]]

mxnet.ndarray.random.generalized_negative_binomial(mu=1, alpha=1, shape=_Null, dtype=_Null, ctx=None, out=None, **kwargs)[source]

Draw random samples from a generalized negative binomial distribution.

Samples are distributed according to a generalized negative binomial distribution parametrized by mu (mean) and alpha (dispersion). alpha is defined as 1/k where k is the failure limit of the number of unsuccessful experiments (generalized to real numbers). Samples will always be returned as a floating point data type.

Parameters:
  • mu (float or NDArray) – Mean of the negative binomial distribution.
  • alpha (float or NDArray) – Alpha (dispersion) parameter of the negative binomial distribution.
  • shape (int or tuple of ints) – The number of samples to draw. If shape is, e.g., (m, n) and mu and alpha are scalars, output shape will be (m, n). If mu and alpha are NDArrays with shape, e.g., (x, y), then output will have shape (x, y, m, n), where m*n samples are drawn for each [mu, alpha) pair.
  • dtype ({'float16','float32', 'float64'}) – Data type of output samples. Default is ‘float32’
  • ctx (Context) – Device context of output. Default is current context. Overridden by mu.context when mu is an NDArray.
  • out (NDArray) – Store output to an existing NDArray.

Examples

>>> mx.nd.random.generalized_negative_binomial(10, 0.5)
[ 19.]

>>> mx.nd.random.generalized_negative_binomial(10, 0.5, shape=(2,))
[ 30.  21.]

>>> mu = mx.nd.array([1,2,3])
>>> alpha = mx.nd.array([0.2,0.4,0.6])
>>> mx.nd.random.generalized_negative_binomial(mu, alpha, shape=2)
[[ 4.  0.]
 [ 3.  2.]
 [ 6.  2.]]

mxnet.ndarray.random.multinomial(data, shape=_Null, get_prob=False, out=None, **kwargs)[source]

Concurrent sampling from multiple multinomial distributions.

Note

The input distribution must be normalized, i.e. data must sum to 1 along its last dimension.

Parameters:
  • data (NDArray) – An n dimensional array whose last dimension has length k, where k is the number of possible outcomes of each multinomial distribution. For example, data with shape (m, n, k) specifies m*n multinomial distributions each with k possible outcomes.
  • shape (int or tuple of ints) – The number of samples to draw from each distribution. If shape is empty one sample will be drawn from each distribution.
  • get_prob (bool) – If true, a second array containing log likelihood of the drawn samples will also be returned. This is usually used for reinforcement learning, where you can provide reward as head gradient w.r.t. this array to estimate gradient.
  • out (NDArray) – Store output to an existing NDArray.

Examples

>>> probs = mx.nd.array([[0, 0.1, 0.2, 0.3, 0.4], [0.4, 0.3, 0.2, 0.1, 0]])
>>> mx.nd.random.multinomial(probs)
[3 1]

>>> mx.nd.random.multinomial(probs, shape=2)
[[4 4]
 [1 2]]

>>> mx.nd.random.multinomial(probs, get_prob=True)
[3 2]

[-1.20397282 -1.60943794]

mxnet.ndarray.random.shuffle(data, **kwargs)[source]

Shuffle the elements randomly.

This shuffles the array along the first axis. The order of the elements in each subarray does not change. For example, if a 2D array is given, the order of the rows randomly changes, but the order of the elements in each row does not change.

Parameters:
  • data (NDArray) – Input data array.
  • out (NDArray) – Array to store the result.

Examples

>>> data = mx.nd.array([[0, 1, 2], [3, 4, 5], [6, 7, 8]])
>>> mx.nd.random.shuffle(data)
[[ 0.  1.  2.]
 [ 6.  7.  8.]
 [ 3.  4.  5.]]

>>> mx.nd.random.shuffle(data)
[[ 3.  4.  5.]
 [ 0.  1.  2.]
 [ 6.  7.  8.]]

Random number interface of MXNet.

mxnet.random.seed(seed_state, ctx='all')[source]

Seeds the random number generators in MXNet.

This affects the behavior of modules in MXNet that uses random number generators, like the dropout operator and NDArray‘s random sampling operators.

Parameters:
  • seed_state (int) – The random number seed.
  • ctx (Context) – The device context of the generator. The default is “all” which means seeding random number generators of all devices.

Notes

Random number generators in MXNet are device specific. mx.random.seed(seed_state) sets the state of each generator using seed_state and the device id. Therefore, random numbers generated from different devices can be different even if they are seeded using the same seed.

To produce identical random number sequences independent of the device id, set optional ctx argument. This produces the same sequence of random numbers independent of the device id, but the sequence can be different on different kind of devices as MXNet’s random number generators for CPU and GPU use different algorithms.

Example

>>> print(mx.nd.random.normal(shape=(2,2)).asnumpy())
[[ 1.36481571 -0.62203991]
 [-1.4962182  -0.08511394]]
>>> print(mx.nd.random.normal(shape=(2,2)).asnumpy())
[[ 1.09544981 -0.20014545]
 [-0.20808885  0.2527658 ]]
# Same results on the same device with the same seed
>>> mx.random.seed(128)
>>> print(mx.nd.random.normal(shape=(2,2)).asnumpy())
[[ 0.47400656 -0.75213492]
 [ 0.20251541  0.95352972]]
>>> mx.random.seed(128)
>>> print(mx.nd.random.normal(shape=(2,2)).asnumpy())
[[ 0.47400656 -0.75213492]
 [ 0.20251541  0.95352972]]
# Different results on gpu(0) and gpu(1) with the same seed
>>> mx.random.seed(128)
>>> print(mx.nd.random.normal(shape=(2,2), ctx=mx.gpu(0)).asnumpy())
[[ 2.5020072 -1.6884501]
 [-0.7931333 -1.4218881]]
>>> mx.random.seed(128)
>>> print(mx.nd.random.normal(shape=(2,2), ctx=mx.gpu(1)).asnumpy())
[[ 0.24336822 -1.664805  ]
 [-1.0223296   1.253198  ]]
# Seeding with `ctx` argument produces identical results on gpu(0) and gpu(1)
>>> mx.random.seed(128, ctx=mx.gpu(0))
>>> print(mx.nd.random.normal(shape=(2,2), ctx=mx.gpu(0)).asnumpy())
[[ 2.5020072 -1.6884501]
 [-0.7931333 -1.4218881]]
>>> mx.random.seed(128, ctx=mx.gpu(1))
>>> print(mx.nd.random.normal(shape=(2,2), ctx=mx.gpu(1)).asnumpy())
[[ 2.5020072 -1.6884501]
 [-0.7931333 -1.4218881]]