Struct AutogradGraph 

pub struct AutogradGraph {
    pub nodes: HashMap<NodeId, AutogradNode>,
    /* private fields */
}

Reverse-mode automatic differentiation graph.

Build a computational graph by calling AutogradGraph::input, AutogradGraph::add, AutogradGraph::mul, etc., then call AutogradGraph::backward on the scalar output node to populate .grad on every node that has requires_grad = true.

Fields

nodes: HashMap<NodeId, AutogradNode>

All nodes keyed by their NodeId.

Implementations

impl AutogradGraph

pub fn new() -> Self

Create an empty graph.

pub fn input(&mut self, value: f64) -> NodeId

Create a leaf input node with requires_grad = true.

pub fn constant(&mut self, value: f64) -> NodeId

Create a constant leaf node with requires_grad = false. Gradients do not flow through constants.

pub fn add(&mut self, lhs: NodeId, rhs: NodeId) -> NodeId

z = lhs + rhs.

pub fn mul(&mut self, lhs: NodeId, rhs: NodeId) -> NodeId

z = lhs * rhs.

pub fn neg(&mut self, input: NodeId) -> NodeId

z = -input.

pub fn exp(&mut self, input: NodeId) -> NodeId

z = e^input.

pub fn ln(&mut self, input: NodeId) -> NodeId

z = ln(input).

pub fn pow(&mut self, base: NodeId, exponent: f64) -> NodeId

z = base^exponent where exponent is a scalar constant.

pub fn backward(&mut self, output_id: NodeId)

Run reverse-mode backpropagation starting from output_id.

Sets nodes[output_id].grad = 1.0, performs a topological sort (reverse post-order DFS), and propagates gradients backwards through the graph. Only nodes with requires_grad = true accumulate gradients.

pub fn grad(&self, node_id: NodeId) -> Option<f64>

Return the gradient of node_id if it exists and has requires_grad = true.

pub fn value(&self, node_id: NodeId) -> Option<f64>

Return the forward-pass value of node_id, or None if not found.

pub fn zero_grad(&mut self)

Reset all accumulated gradients to 0.0.

Trait Implementations

impl Default for AutogradGraph

fn default() -> Self

Returns the “default value” for a type.