View Source Nx.Random (Nx v0.5.3)
Pseudo-random number generators.
Unlike the stateful pseudo-random number generators (PRNGs)
that users of most programming languages and numerical libraries
may be accustomed to, Nx random functions require an explicit
PRNG key to be passed as a first argument. That key is defined by
an Nx.Tensor composed of 2 unsigned 32-bit integers, usually
generated by the Nx.Random.key/1 function:
iex> Nx.Random.key(12)
#Nx.Tensor<
u32[2]
[0, 12]
>This key can then be used in any of Nx’s random number generation routines:
iex> key = Nx.Random.key(12)
iex> {uniform, _new_key} = Nx.Random.uniform(key)
iex> uniform
#Nx.Tensor<
f32
0.7691127061843872
>Now, when generating a new random number, you pass the new_key
to get a different number.
The function in this module also have a *_split variant, which
is used when the key has been split before hand.
design-and-context
Design and Context
In short, Nx's PRNGs are based on a Threefry counter PRNG associated to a functional array-oriented splitting model. To summarize, among other requirements, Nx's PRNG aims to:
Ensure reproducibility
Parallelize well, both in terms of vectorization (generating array values) and multi-replica, multi-core computation. In particular it should not use sequencing constraints between random function calls.
Link to this section Summary
Functions
Generates random samples from a tensor.
Generates random samples from a tensor with specified probabilities.
Folds in new data to a PRNG key.
Sample Gumbel random values with given shape and float dtype.
Same as gumbel/2, but assumes the key has been split beforehand.
Create a pseudo-random number generator (PRNG) key given an integer seed.
Shortcut for normal(key, 0.0, 1.0, opts).
Returns a normal distribution with the given mean and standard_deviation.
Same as normal/4 but assumes the key has already been split.
Sample uniform random integer values in [min_value, max_value).
Same as randint/4 but assumes the key has already been split.
Randomly shuffles tensor elements along an axis.
Splits a PRNG key into num new keys by adding a leading axis.
Shortcut for uniform(key, 0.0, 1.0, opts).
Sample uniform float values in [min_val, max_val).
Same as uniform/4 but assumes the key has already been split.
Link to this section Functions
Generates random samples from a tensor.
options
Options
:samples- The number of samples to take:axis- The axis along which to take samples. Ifnil, the tensor is flattened beforehand.:replace- a boolean that specifies if samples will be taken with or without replacement. Defaults totrue.
examples
Examples
iex> k = Nx.Random.key(1)
iex> t = Nx.iota({4, 3})
iex> {result, _key} = Nx.Random.choice(k, t, samples: 4, axis: 0)
iex> result
#Nx.Tensor<
s64[4][3]
[
[6, 7, 8],
[3, 4, 5],
[6, 7, 8],
[3, 4, 5]
]
>
iex> {result, _key} = Nx.Random.choice(k, t, samples: 4, axis: 0, replace: false)
iex> result
#Nx.Tensor<
s64[4][3]
[
[3, 4, 5],
[9, 10, 11],
[6, 7, 8],
[0, 1, 2]
]
>If no axis is specified, the tensor is flattened:
iex> k = Nx.Random.key(2)
iex> t = Nx.iota({3, 2})
iex> {result, _key} = Nx.Random.choice(k, t)
iex> result
#Nx.Tensor<
s64[1]
[3]
>
iex> {result, _key} = Nx.Random.choice(k, t, samples: 6, replace: false)
iex> result
#Nx.Tensor<
s64[6]
[2, 0, 4, 5, 1, 3]
>Generates random samples from a tensor with specified probabilities.
The probabilities tensor must have the same size as the axis along which the samples are being taken. If no axis is given, the size must be equal to the input tensor's size.
options
Options
:samples- The number of samples to take:axis- The axis along which to take samples. Ifnil, the tensor is flattened beforehand.:replace- a boolean that specifies if samples will be taken with or without replacement. Defaults totrue.
examples
Examples
iex> k = Nx.Random.key(1)
iex> t = Nx.iota({4, 3})
iex> p = Nx.tensor([0.1, 0.7, 0.2])
iex> {result, _key} = Nx.Random.choice(k, t, p, samples: 3, axis: 1)
iex> result
#Nx.Tensor<
s64[4][3]
[
[1, 0, 1],
[4, 3, 4],
[7, 6, 7],
[10, 9, 10]
]
>
iex> {result, _key} = Nx.Random.choice(k, t, p, samples: 3, axis: 1, replace: false)
iex> result
#Nx.Tensor<
s64[4][3]
[
[1, 2, 0],
[4, 5, 3],
[7, 8, 6],
[10, 11, 9]
]
>If no axis is specified, the tensor is flattened. Notice that in the first case we get a higher occurence of the entries with bigger probabilities, while in the second case, without replacements, we get those samples first.
iex> k = Nx.Random.key(2)
iex> t = Nx.iota({2, 3})
iex> p = Nx.tensor([0.01, 0.1, 0.19, 0.6, 0.05, 0.05])
iex> {result, _key} = Nx.Random.choice(k, t, p)
iex> result
#Nx.Tensor<
s64[1]
[3]
>
iex> {result, _key} = Nx.Random.choice(k, t, p, samples: 6)
iex> result
#Nx.Tensor<
s64[6]
[3, 3, 3, 0, 3, 3]
>
iex> {result, _key} = Nx.Random.choice(k, t, p, samples: 6, replace: false)
iex> result
#Nx.Tensor<
s64[6]
[3, 1, 2, 5, 4, 0]
>Folds in new data to a PRNG key.
examples
Examples
iex> key = Nx.Random.key(42)
iex> Nx.Random.fold_in(key, 99)
#Nx.Tensor<
u32[2]
[2015327502, 1351855566]
>
iex> key = Nx.Random.key(42)
iex> Nx.Random.fold_in(key, 1234)
#Nx.Tensor<
u32[2]
[1356445167, 2917756949]
>
iex> key = Nx.Random.key(42)
iex> Nx.Random.fold_in(key, Nx.tensor([[1, 99], [1234, 13]]))
#Nx.Tensor<
u32[2][2][2]
[
[
[64467757, 2916123636],
[2015327502, 1351855566]
],
[
[1356445167, 2917756949],
[3514951389, 229662949]
]
]
>Sample Gumbel random values with given shape and float dtype.
options
Options
:shape- the shape of the output tensor containing the random samples. Defaults to{}:type- the floating-point output type. Defaults to{:f, 32}
examples
Examples
iex> {result, _key} = Nx.Random.gumbel(Nx.Random.key(1))
iex> result
#Nx.Tensor<
f32
-0.7294610142707825
>
iex> {result, _key} = Nx.Random.gumbel(Nx.Random.key(1), shape: {2, 3})
iex> result
#Nx.Tensor<
f32[2][3]
[
[0.6247938275337219, -0.21740718185901642, 0.7678327560424805],
[0.7778404355049133, 4.0895304679870605, 0.3029090166091919]
]
>Same as gumbel/2, but assumes the key has been split beforehand.
Create a pseudo-random number generator (PRNG) key given an integer seed.
examples
Examples
iex> Nx.Random.key(12)
#Nx.Tensor<
u32[2]
[0, 12]
>
iex> Nx.Random.key(999999999999)
#Nx.Tensor<
u32[2]
[232, 3567587327]
>Shortcut for normal(key, 0.0, 1.0, opts).
Returns a normal distribution with the given mean and standard_deviation.
options
Options
:type- a float or complex type for the returned tensor:shape- shape of the returned tensor:names- the names of the returned tensor
examples
Examples
iex> key = Nx.Random.key(42)
iex> {normal, _new_key} = Nx.Random.normal(key)
iex> normal
#Nx.Tensor<
f32
1.3694695234298706
>
iex> key = Nx.Random.key(42)
iex> {normal, _new_key} = Nx.Random.normal(key, 0, 1, shape: {3, 2}, type: :f16)
iex> normal
#Nx.Tensor<
f16[3][2]
[
[-0.32568359375, -0.77197265625],
[0.39208984375, 0.5341796875],
[0.270751953125, -2.080078125]
]
>
iex> key = Nx.Random.key(42)
iex> {normal, _new_key} = Nx.Random.normal(key, 0, 1, shape: {2, 2}, type: :c64)
iex> normal
#Nx.Tensor<
c64[2][2]
[
[-0.7632761001586914+0.8661127686500549i, -0.14282889664173126-0.7384796142578125i],
[0.678461492061615+0.4118310809135437i, -2.269538402557373-0.3689095079898834i]
]
>
iex> key = Nx.Random.key(1337)
iex> {normal, _new_key} = Nx.Random.normal(key, 10, 5, shape: {1_000})
iex> Nx.mean(normal)
#Nx.Tensor<
f32
9.70022201538086
>
iex> Nx.standard_deviation(normal)
#Nx.Tensor<
f32
5.051416397094727
>Same as normal/4 but assumes the key has already been split.
Sample uniform random integer values in [min_value, max_value).
options
Options
:type- the integer type for the returned tensor:shape- shape of the returned tensor:names- the names of the returned tensor
examples
Examples
iex> key = Nx.Random.key(1701)
iex> {randint, _new_key} = Nx.Random.randint(key, 1, 100)
iex> randint
#Nx.Tensor<
s64
66
>
iex> key = Nx.Random.key(1701)
iex> {randint, _new_key} = Nx.Random.randint(key, 1, 100, shape: {3, 2}, type: :u32)
iex> randint
#Nx.Tensor<
u32[3][2]
[
[9, 20],
[19, 6],
[71, 15]
]
>Same as randint/4 but assumes the key has already been split.
Randomly shuffles tensor elements along an axis.
options
Options
:axis- the axis along which to shuffle. Defaults to0:independent- a boolean that indicates wether the permutations are independent along the given axis. Defaults tofalse
examples
Examples
iex> key = Nx.Random.key(42)
iex> {shuffled, _new_key} = Nx.Random.shuffle(key, Nx.iota({3, 4}, axis: 0))
iex> shuffled
#Nx.Tensor<
s64[3][4]
[
[2, 2, 2, 2],
[0, 0, 0, 0],
[1, 1, 1, 1]
]
>
iex> key = Nx.Random.key(10)
iex> {shuffled, _new_key} = Nx.Random.shuffle(key, Nx.iota({3, 4}, axis: 1), independent: true, axis: 1)
iex> shuffled
#Nx.Tensor<
s64[3][4]
[
[2, 1, 3, 0],
[3, 0, 1, 2],
[2, 3, 0, 1]
]
>Splits a PRNG key into num new keys by adding a leading axis.
examples
Examples
iex> key = Nx.Random.key(1701)
iex> Nx.Random.split(key)
#Nx.Tensor<
u32[2][2]
[
[56197195, 1801093307],
[961309823, 1704866707]
]
>
iex> key = Nx.Random.key(999999999999)
iex> Nx.Random.split(key, parts: 4)
#Nx.Tensor<
u32[4][2]
[
[3959978897, 4079927650],
[3769699049, 3585271160],
[3182829676, 333122445],
[3185556048, 1258545461]
]
>Shortcut for uniform(key, 0.0, 1.0, opts).
Sample uniform float values in [min_val, max_val).
options
Options
:type- a float type for the returned tensor:shape- shape of the returned tensor:names- the names of the returned tensor
examples
Examples
iex> key = Nx.Random.key(1701)
iex> {uniform, _new_key} = Nx.Random.uniform(key)
iex> uniform
#Nx.Tensor<
f32
0.9728643894195557
>
iex> key = Nx.Random.key(1701)
iex> {uniform, _new_key} = Nx.Random.uniform(key, shape: {3, 2}, type: :f16)
iex> uniform
#Nx.Tensor<
f16[3][2]
[
[0.75390625, 0.6484375],
[0.7294921875, 0.21484375],
[0.09765625, 0.0693359375]
]
>
iex> key = Nx.Random.key(1701)
iex> {uniform, _new_key} = Nx.Random.uniform(key, shape: {2, 2}, type: :c64)
iex> uniform
#Nx.Tensor<
c64[2][2]
[
[0.18404805660247803+0.6546461582183838i, 0.5525915622711182+0.11568140983581543i],
[0.6074584722518921+0.8104375600814819i, 0.247686505317688+0.21975469589233398i]
]
>Same as uniform/4 but assumes the key has already been split.