Struct Value 

pub struct Value<V: Differentiable> { /* private fields */ }

Public value handle for reverse-mode AD.

Value exposes a torch-like surface while keeping graph ownership hidden. Internally it carries either a leaf gradient sink or an edge into a reverse graph.

Implementations

impl<V: Differentiable + 'static> Value<V>

pub fn new(primal: V) -> Self

Create a detached value.

pub fn primal(&self) -> &V

Borrow the primal value.

pub fn requires_grad(&self) -> bool

Return whether this value participates in reverse-mode AD.

pub fn with_requires_grad(self, enabled: bool) -> Self

Return a new value handle with gradient tracking enabled or disabled.

pub fn grad(&self) -> AdResult<Option<V::Tangent>>

where V::Tangent: Clone,

Read the accumulated leaf gradient, if available.

pub fn zero_grad(&self) -> AdResult<()>

Clear the accumulated leaf gradient.

pub fn backward(&self) -> AdResult<()>

where V::Tangent: Clone,

Run reverse-mode backward with the default scalar cotangent seed.

pub fn backward_with_seed(&self, seed: V::Tangent) -> AdResult<()>

where V::Tangent: Clone,

Run reverse-mode backward with an explicit cotangent seed.

pub fn grad_wrt_with_seed( &self, seed: V::Tangent, wrt: &[&Self], ) -> AdResult<Vec<Option<V::Tangent>>>

where V::Tangent: Clone,

Compute functional gradients with respect to the requested inputs.

pub fn shares_reverse_graph(&self, other: &Self) -> bool

Return whether self and other share any reachable reverse graph.