Struct zyx_core::runtime::Runtime

pub struct Runtime<R: RuntimeBackend> { /* private fields */ }

Runtime with autograd engine. This runtime uses Node enum as representation of tensors.

Implementations

impl<R: RuntimeBackend> Runtime<R>

pub fn new(runtime_backend: R) -> Self

Initialize new runtime.

pub fn randn(&mut self, shape: Shape, dtype: DType) -> Result<Id, ZyxError>

Create tensor initialized from normal distribution.

pub fn uniform<T: Scalar>( &mut self, shape: Shape, range: Range<T> ) -> Result<Id, ZyxError>

Create uniform tensor from range low..high

pub fn shape(&self, x: Id) -> &Shape

Get shape of tensor x

pub fn dtype(&self, x: Id) -> DType

Get dtype of tensor x

pub fn load<T: Scalar>(&mut self, x: Id) -> Result<Vec<T>, ZyxError>

Load tensor x

pub fn store<T: Scalar, IT>(&mut self, iter: IT) -> Result<Id, ZyxError>

where IT: IntoIterator<Item = T>, IT::IntoIter: ExactSizeIterator,

Store iterator into runtime as tensor

pub fn push(&mut self, node: Node) -> Result<Id, ZyxError>

Push new Node into the graph creating new tensor. This function does ZERO verification that the node is correct, but it optimizes out useless operations (like reshaping to the same shape)

pub fn release(&mut self, x: Id) -> Result<(), ZyxError>

Decrease reference count of x. If x’s reference count reaches zero, this function will delete x and release all of it’s predecessors in the graph.

pub fn retain(&mut self, x: Id)

Increase reference count of tensor x.

pub fn debug_graph(&self)

Debug print all nodes

pub fn evaluate(&mut self, nodes: BTreeSet<Id>) -> Result<(), ZyxError>

Evaluate specified nodes.

pub fn plot_graph_dot(&self, ids: &[Id]) -> String

Plot dot graph in dot format between given nodes

pub fn backward( &mut self, x: Id, sources: &BTreeSet<Id> ) -> Result<BTreeMap<Id, Id>, ZyxError>

Common autograd engine, currently used by all backends.