backward method

void backward({ double startingGrad = 1.0, })

Starting point of gradient calculation. The network of Nodes and Tensors traverses in topological order over all Tensors affecting this Tensor adding a given gradient, calculated via their creator. The value of startingGrad affects the first gradient given into the backwards process and is set to 1.0 by default.

Implementation

void backward({double startingGrad = 1.0}) {
  for (int i = 0; i < grad.length; i = i + 1) {
    grad[i] = startingGrad;
  }

  if (creator == null) return;

  List<Node> topo = [];
  Set<Node> visited = {};

  void buildTopo(Node? node) {
    if (node == null || visited.contains(node)) return;
    visited.add(node);
    for (int i = 0; i < node.inputs.length; i = i + 1) {
      Tensor inputTensor = node.inputs[i];
      if (inputTensor.creator != null) {
        buildTopo(inputTensor.creator);
      }
    }
    topo.add(node);
  }

  buildTopo(creator);

  for (int i = topo.length - 1; i >= 0; i = i - 1) {
    topo[i].backwardFn();
  }
}