tenzr::random¶

Classes¶

Functions¶

Functions Documentation¶

function DefaultGenerator¶

Generator & DefaultGenerator()

Return:

• The process-wide default generator.
• Reference to the singleton Generator.

Get the process-wide default generator.

function Key¶

TENZR_APIPRNGKey Key(
    uint64_t seed
)

Parameters:

• seed Seed value.
• seed_val Seed value.

Return:

• The derived key.
• The derived key.

Create a PRNGKey from a 64-bit integer seed. Deterministically derive a PRNG key from a 64-bit seed.

function Threefry2x32¶

TENZR_API std::pair< uint32_t, uint32_t > Threefry2x32(
    uint32_t k0,
    uint32_t k1,
    uint32_t c0,
    uint32_t c1
)

Parameters:

• k0 Key high 32 bits.
• k1 Key low 32 bits.
• c0 Counter high 32 bits.
• c1 Counter low 32 bits.
• k0 First key word.
• k1 Second key word.
• c0 First counter word.
• c1 Second counter word.

Return:

• The (high, low) hashed word.
• The two generated 32-bit words.

Evaluate the 20-round Threefry 2x32 counter-based PRNG hash. Apply the 20-round Threefry 2x32 counter-based PRNG to a key and counter pair.

function Split¶

TENZR_APISplitKeys Split(
    const PRNGKey & k
)

Parameters:

• k Key to split.
• k Key to split.

Return:

• The pair of derived keys.
• Pair of derived keys.

Split a key into two fresh independent keys. Split a key into two independent keys.

function Split¶

TENZR_API std::vector< PRNGKey > Split(
    const PRNGKey & k,
    size_t num
)

Parameters:

• k Key to split.
• num Number of keys to produce.
• k Key to split.
• num Number of keys to derive; must be positive.

Return:

• The derived keys.
• The derived keys.

Split a key into an arbitrary number of fresh independent sub-keys. Split a key into num independent keys.

function FoldIn¶

TENZR_APIPRNGKey FoldIn(
    const PRNGKey & k,
    uint32_t data
)

Parameters:

• k Key to fold into.
• data 32-bit data word.
• k Base key.
• data Data word mixed into the counter.

Return:

• The folded key.
• The derived key.

Deterministically fold arbitrary 32-bit integer data into an existing key. Fold integer data into a key to derive a new key.

function ThreefryRandomBits¶

TENZR_API std::vector< uint32_t > ThreefryRandomBits(
    const PRNGKey & k,
    size_t count
)

Parameters:

• k Key to draw from.
• count Number of words to produce.
• k Key to draw from.
• count Number of words needed.

Return:

• The generated words.
• The generated words.

Generate count random 32-bit unsigned words from a key using Threefry2x32. Generate count pseudorandom 32-bit words from a key.

function Uniform¶

TENZR_APIArray Uniform(
    const PRNGKey & key,
    const Shape & shape,
    float low =0.0f,
    float high =1.0f,
    DType dtype =DType::kFloat32
)

Parameters:

• key PRNG key.
• shape Output shape.
• low Inclusive lower bound.
• high Exclusive upper bound.
• dtype Output dtype.
• key Key to draw from.
• shape Output shape.
• low Inclusive lower bound.
• high Exclusive upper bound.
• dtype Result dtype (kFloat32 or Float64).

Return:

• Sampled array.
• The sampled array.

Sample uniform random values in [low, high). Draw a uniform sample over [low, high) using a functional key.

function Normal¶

TENZR_APIArray Normal(
    const PRNGKey & key,
    const Shape & shape,
    float mean =0.0f,
    float stddev =1.0f,
    DType dtype =DType::kFloat32
)

Parameters:

• key PRNG key.
• shape Output shape.
• mean Distribution mean.
• stddev Distribution standard deviation.
• dtype Output dtype.
• key Key to draw from.
• shape Output shape.
• mean Distribution mean.
• stddev Distribution standard deviation, which must be positive.
• dtype Result dtype (kFloat32 or Float64).

Return:

• Sampled array.
• The sampled array.

Sample normal (Gaussian) values with the given mean and standard deviation. Draw a normal sample via Box-Muller using a functional key.

function Bernoulli¶

TENZR_APIArray Bernoulli(
    const PRNGKey & key,
    const Shape & shape,
    float p =0.5f
)

Parameters:

• key PRNG key.
• shape Output shape.
• p Probability of true.
• key Key to draw from.
• shape Output shape.
• p Success probability, in [0, 1].

Return:

• Boolean array.
• The sampled boolean array.

Sample booleans that are true with probability p. Draw a boolean Bernoulli sample using a functional key.

function Rademacher¶

TENZR_APIArray Rademacher(
    const PRNGKey & key,
    const Shape & shape,
    DType dtype =DType::kFloat32
)

Parameters:

• key PRNG key.
• shape Output shape.
• dtype Output dtype.
• key Key to draw from.
• shape Output shape.
• dtype Result dtype (kFloat32 or Float64).

Return:

• Sampled array.
• The sampled array.

Sample Rademacher values (+1 or -1 with equal probability). Draw a Rademacher sample of +1/-1 using a functional key.

function Categorical¶

TENZR_APIArray Categorical(
    const PRNGKey & key,
    const Array & logits,
    const Shape & shape
)

Parameters:

• key PRNG key.
• logits Unnormalized class scores.
• shape Output shape.
• key Key to draw from.
• logits Unnormalized class scores; the last dimension is the class axis.
• shape Output shape; its element count is the number of samples.

Return:

• Sampled class indices.
• Int32 array of sampled class indices.

Draw categorical samples from log-probabilities via the Gumbel-max trick. Draw class indices from a categorical distribution over logits.

function GlorotUniform¶

TENZR_APIArray GlorotUniform(
    const PRNGKey & key,
    const Shape & shape,
    int64_t fan_in,
    int64_t fan_out,
    DType dtype =DType::kFloat32
)

Parameters:

• key PRNG key.
• shape Output shape.
• fan_in Number of input units.
• fan_out Number of output units.
• dtype Output dtype.
• key Key to draw from.
• shape Output shape.
• fan_in Input fan size.
• fan_out Output fan size.
• dtype Result dtype.

Return:

• Initialized array.
• The initialized array.

Glorot/Xavier uniform initialization. Glorot (Xavier) uniform initialization using a functional key.

function GlorotNormal¶

TENZR_APIArray GlorotNormal(
    const PRNGKey & key,
    const Shape & shape,
    int64_t fan_in,
    int64_t fan_out,
    DType dtype =DType::kFloat32
)

Parameters:

• key PRNG key.
• shape Output shape.
• fan_in Number of input units.
• fan_out Number of output units.
• dtype Output dtype.
• key Key to draw from.
• shape Output shape.
• fan_in Input fan size.
• fan_out Output fan size.
• dtype Result dtype.

Return:

• Initialized array.
• The initialized array.

Glorot/Xavier normal initialization. Glorot (Xavier) normal initialization using a functional key.

function KaimingUniform¶

TENZR_APIArray KaimingUniform(
    const PRNGKey & key,
    const Shape & shape,
    int64_t fan_in,
    DType dtype =DType::kFloat32
)

Parameters:

• key PRNG key.
• shape Output shape.
• fan_in Number of input units.
• dtype Output dtype.
• key Key to draw from.
• shape Output shape.
• fan_in Input fan size.
• dtype Result dtype.

Return:

• Initialized array.
• The initialized array.

Kaiming/He uniform initialization. Kaiming (He) uniform initialization using a functional key.

function KaimingNormal¶

TENZR_APIArray KaimingNormal(
    const PRNGKey & key,
    const Shape & shape,
    int64_t fan_in,
    DType dtype =DType::kFloat32
)

Parameters:

• key PRNG key.
• shape Output shape.
• fan_in Number of input units.
• dtype Output dtype.
• key Key to draw from.
• shape Output shape.
• fan_in Input fan size.
• dtype Result dtype.

Return:

• Initialized array.
• The initialized array.

Kaiming/He normal initialization. Kaiming (He) normal initialization using a functional key.

function Seed¶

void Seed(
    uint64_t s
)

Parameters:

• s New seed.
• s New seed value.

Reseed the default generator. Reseed the default generator deterministically.

function Seed¶

void Seed()

Reseed the default generator from a non-deterministic source.

Reseed the default generator from system randomness.

function Uniform¶

Tensor Uniform(
    Shape shape,
    float low =0.0f,
    float high =1.0f,
    DType dtype =DType::kFloat32,
    Generator & gen =DefaultGenerator()
)

Parameters:

• shape Output shape.
• low Inclusive lower bound.
• high Exclusive upper bound.
• dtype Output dtype.
• gen Generator to draw from.
• shape Output shape.
• low Inclusive lower bound.
• high Exclusive upper bound.
• dtype Result dtype (kFloat32 or Float64).
• gen Stateful generator to draw from.

Return:

• Sampled array.
• The sampled tensor.

Sample uniform values in [low, high) from gen. Draw a uniform tensor from a stateful generator.

function Normal¶

Tensor Normal(
    Shape shape,
    float mean =0.0f,
    float stddev =1.0f,
    DType dtype =DType::kFloat32,
    Generator & gen =DefaultGenerator()
)

Parameters:

• shape Output shape.
• mean Distribution mean.
• stddev Distribution standard deviation.
• dtype Output dtype.
• gen Generator to draw from.
• shape Output shape.
• mean Distribution mean.
• stddev Distribution standard deviation, which must be positive.
• dtype Result dtype (kFloat32 or Float64).
• gen Stateful generator to draw from.

Return:

• Sampled array.
• The sampled tensor.

Sample normal values from gen. Draw a normal tensor from a stateful generator.

function Bernoulli¶

Tensor Bernoulli(
    Shape shape,
    float p =0.5f,
    Generator & gen =DefaultGenerator()
)

Parameters:

• shape Output shape.
• p Probability of true.
• gen Generator to draw from.
• shape Output shape.
• p Success probability, in [0, 1].
• gen Stateful generator to draw from.

Return:

• Boolean array.
• The sampled 0.0/1.0 tensor.

Sample booleans from gen. Draw a kFloat32 Bernoulli mask from a stateful generator.

function GlorotUniform¶

Tensor GlorotUniform(
    Shape shape,
    int64_t fan_in,
    int64_t fan_out,
    DType dtype,
    Generator & gen =DefaultGenerator()
)

Glorot uniform initialization with an explicit dtype.

Parameters:

• shape Output shape.
• fan_in Input fan size.
• fan_out Output fan size.
• dtype Result dtype.
• gen Stateful generator to draw from.

Return: The initialized tensor.

Glorot uniform initialization from a stateful generator.

function GlorotUniform¶

Tensor GlorotUniform(
    Shape shape,
    int64_t fan_in,
    int64_t fan_out,
    Generator & gen =DefaultGenerator()
)

Glorot uniform initialization.

Parameters:

• shape Output shape.
• fan_in Input fan size.
• fan_out Output fan size.
• gen Stateful generator to draw from.

Return: The initialized tensor.

Glorot uniform initialization with a kFloat32 default dtype.

function GlorotNormal¶

Tensor GlorotNormal(
    Shape shape,
    int64_t fan_in,
    int64_t fan_out,
    DType dtype,
    Generator & gen =DefaultGenerator()
)

Glorot normal initialization with an explicit dtype.

Parameters:

• shape Output shape.
• fan_in Input fan size.
• fan_out Output fan size.
• dtype Result dtype.
• gen Stateful generator to draw from.

Return: The initialized tensor.

Glorot normal initialization from a stateful generator.

function GlorotNormal¶

Tensor GlorotNormal(
    Shape shape,
    int64_t fan_in,
    int64_t fan_out,
    Generator & gen =DefaultGenerator()
)

Glorot normal initialization.

Parameters:

• shape Output shape.
• fan_in Input fan size.
• fan_out Output fan size.
• gen Stateful generator to draw from.

Return: The initialized tensor.

Glorot normal initialization with a kFloat32 default dtype.

function KaimingUniform¶

Tensor KaimingUniform(
    Shape shape,
    int64_t fan_in,
    DType dtype,
    Generator & gen =DefaultGenerator()
)

Kaiming uniform initialization with an explicit dtype.

Parameters:

• shape Output shape.
• fan_in Input fan size.
• dtype Result dtype.
• gen Stateful generator to draw from.

Return: The initialized tensor.

Kaiming uniform initialization from a stateful generator.

function KaimingUniform¶

Tensor KaimingUniform(
    Shape shape,
    int64_t fan_in,
    Generator & gen =DefaultGenerator()
)

Kaiming uniform initialization.

Parameters:

• shape Output shape.
• fan_in Input fan size.
• gen Stateful generator to draw from.

Return: The initialized tensor.

Kaiming uniform initialization with a kFloat32 default dtype.

function KaimingNormal¶

Tensor KaimingNormal(
    Shape shape,
    int64_t fan_in,
    DType dtype,
    Generator & gen =DefaultGenerator()
)

Kaiming normal initialization with an explicit dtype.

Parameters:

• shape Output shape.
• fan_in Input fan size.
• dtype Result dtype.
• gen Stateful generator to draw from.

Return: The initialized tensor.

Kaiming normal initialization from a stateful generator.

function KaimingNormal¶

Tensor KaimingNormal(
    Shape shape,
    int64_t fan_in,
    Generator & gen =DefaultGenerator()
)

Kaiming normal initialization.

Parameters:

• shape Output shape.
• fan_in Input fan size.
• gen Stateful generator to draw from.

Return: The initialized tensor.

Kaiming normal initialization with a kFloat32 default dtype.

function Rand¶

inline Tensor Rand(
    Shape shape,
    Generator & gen =DefaultGenerator()
)

Parameters:

• shape Output shape.
• gen Generator to draw from.

Return: Sampled array.

Sample uniform values in [0, 1).

function Rand¶

inline Tensor Rand(
    Shape shape,
    float low,
    float high,
    Generator & gen =DefaultGenerator()
)

Parameters:

• shape Output shape.
• low Inclusive lower bound.
• high Exclusive upper bound.
• gen Generator to draw from.

Return: Sampled array.

Sample uniform values in [low, high).

function Randn¶

inline Tensor Randn(
    Shape shape,
    Generator & gen =DefaultGenerator()
)

Parameters:

• shape Output shape.
• gen Generator to draw from.

Return: Sampled array.

Sample standard normal values.

function Randn¶

inline Tensor Randn(
    Shape shape,
    float mean,
    float stddev,
    Generator & gen =DefaultGenerator()
)

Parameters:

• shape Output shape.
• mean Distribution mean.
• stddev Distribution standard deviation.
• gen Generator to draw from.

Return: Sampled array.

Sample normal values.