Gradient

Overview

• namespace: Rindow\NeuralNetworks\Builder
• classname: Gradient

Creation of objects with automatic differentiation function

gradient utilities

• Variable : Create variable
• GradientTape : Tape for recording calculation graphs
• GraphFunction : Creating a differentiable function
• isUndetermined: Check variable.

Differentiable functions

• abs
• add
• sub
• mul
• div
• square
• sqrt
• exp
• log
• log1p
• matmul
• reduceMean
• reduceSum
• reduceMax
• concat
• softmax
• logSoftmax
• masking
• gather
• expandDims
• squeeze
• minimum
• maximum
• randomNormal
• randomCategorical
• tanh
• relu
• sigmoid
• where
• l2norm
• etc.

Examples

use Rindow\Math\Matrix\MatrixOperator;
use Rindow\NeuralNetworks\Builder\NeuralNetworks;
$mo = new MatrixOperator();
$nn = new NeuralNetworks($mo);
$g = $nn->gradient();
$a = $g->Variable([1,2]);
$b = $g->Variable([2,3]);
$c = $nn->with($tape=$g->GradientTape(),function() use ($g,$a,$b) {
    return $g->mul($a,$b);
});
[$da,$db] = $tape->gradient($c,[$a,$b]);
echo $mo->toString($c)."\n";
echo $mo->toString($da)."\n";
echo $mo->toString($db)."\n";

# [2,6]
# [2,3]
# [1,2]