Struct Var 

pub struct Var(/* private fields */);

Implementations

impl Var

pub fn leaf(t: Tensor) -> Var

A differentiable leaf (a parameter or input we want gradients for).

pub fn value(&self) -> &Tensor

pub fn grad(&self) -> Option<Tensor>

pub fn backward(&self)

Backpropagate from this (scalar-ish) output: seed grad = ones, walk reverse-topo.

pub fn add(&self, o: &Var) -> Var

pub fn sub(&self, o: &Var) -> Var

pub fn mul(&self, o: &Var) -> Var

pub fn matmul(&self, o: &Var) -> Var

Matmul (last two dims; equal or absent batch dims).

pub fn relu(&self) -> Var

pub fn neg(&self) -> Var

pub fn transpose(&self, a: usize, b: usize) -> Var

Transpose two dims; gradient transposes back.

pub fn div(&self, o: &Var) -> Var

pub fn exp(&self) -> Var

pub fn log(&self) -> Var

pub fn sum(&self, axes: &[usize]) -> Var

Sum over axes (keepdim); gradient broadcasts back to the input shape.

pub fn mean(&self, axes: &[usize]) -> Var

pub fn detach(&self) -> Var

A non-differentiable copy (stop-gradient) — for the detached max in a stable softmax.

pub fn softmax(&self, axis: usize) -> Var

Numerically-stable softmax over axis, fully differentiable (built from primitives).

pub fn sum_all(&self) -> Var

pub fn mean_all(&self) -> Var

Trait Implementations

impl Clone for Var

fn clone(&self) -> Var

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.