randomCategorical

• namespace: Rindow\NeuralNetworks\Gradient\Func
• classname: RandomCategorical

Random categorical sampling function.

Samples category indices from logits. Gradients are not propagated to the input.

Methods

randomCategorical

$g->randomCategorical(
    Variable|NDArray $logits,
    ?array $batchShape=null,
    ?bool $softmax=null,
    ?int $dtype=null,
    ?int $seed=null,
) : Variable

Create and execute the function in the builder method

Arguments

• logits: A Variable or NDArray of unnormalized log probabilities. The last dimension is the number of categories. Implicitly create Variable for NDArray.

Options

• batchShape: Shape of the batch to sample. If specified, the input must be one-dimensional.
• softmax: If true, apply softmax to the logits before sampling. The default is true.
• dtype: Data type of the output indices. The default is int32.
• seed: Random seed.

use Rindow\Math\Matrix\MatrixOperator;
use Rindow\NeuralNetworks\Builder\NeuralNetworks;
$mo = new MatrixOperator();
$nn = new NeuralNetworks($mo);
$g = $nn->gradient();
$logits = $g->Variable([[1,2,3],[1,1,1]]);
$y = $g->randomCategorical($logits,seed:0);
echo $mo->toString($y)."\n";

# shape: [2]