Struct Variable 

pub struct Variable { /* private fields */ }

A tensor with automatic differentiation support.

Variable wraps a Tensor and tracks operations performed on it to enable automatic gradient computation. When requires_grad is true, all operations are recorded in a computational graph.

Implementations

impl Variable

pub fn new(data: Tensor<f32>, requires_grad: bool) -> Variable

Creates a new variable from a tensor.

Arguments

• data - The tensor data
• requires_grad - Whether to track gradients for this variable

pub fn from_tensor(data: Tensor<f32>) -> Variable

Creates a variable that doesn’t require gradients.

pub fn from_operation( data: Tensor<f32>, grad_fn: GradFn, requires_grad: bool, ) -> Variable

Creates a new variable from an operation result with an attached gradient function.

This connects the variable to the computational graph, allowing gradients to flow backward through the operation that produced this variable.

pub fn data(&self) -> Tensor<f32>

Returns a reference to the underlying tensor data.

pub fn shape(&self) -> Vec<usize>

Returns the shape of the tensor.

pub fn ndim(&self) -> usize

Returns the number of dimensions.

pub fn numel(&self) -> usize

Returns the total number of elements.

pub fn requires_grad(&self) -> bool

Returns whether this variable requires gradients.

pub fn is_leaf(&self) -> bool

Returns whether this is a leaf variable.

pub fn grad(&self) -> Option<Tensor<f32>>

Returns the gradient of this variable.

Only available for leaf variables after backward() has been called.

pub fn grad_fn(&self) -> Option<&GradFn>

Returns the gradient function.

pub fn set_grad(&self, grad: Tensor<f32>)

Sets the gradient (used during backward pass).

pub fn accumulate_grad(&self, grad: &Tensor<f32>)

Accumulates gradient (adds to existing gradient).

pub fn zero_grad(&self)

Clears the gradient.

pub fn detach(&self) -> Variable

Detaches this variable from the computation graph.

Returns a new variable with the same data but no gradient history.

pub fn requires_grad_(self, requires_grad: bool) -> Variable

Returns a new variable with requires_grad set.

pub fn backward(&self)

Computes gradients via backpropagation.

This should only be called on scalar (single-element) tensors, typically the loss value.

pub fn add_var(&self, other: &Variable) -> Variable

Element-wise addition.

pub fn sub_var(&self, other: &Variable) -> Variable

Element-wise subtraction.

pub fn mul_var(&self, other: &Variable) -> Variable

Element-wise multiplication.

pub fn div_var(&self, other: &Variable) -> Variable

Element-wise division.

pub fn neg_var(&self) -> Variable

Negation.

pub fn matmul(&self, other: &Variable) -> Variable

Matrix multiplication.

pub fn pow(&self, exponent: f32) -> Variable

Power operation.

pub fn relu(&self) -> Variable

ReLU activation.

pub fn sigmoid(&self) -> Variable

Sigmoid activation.

pub fn tanh(&self) -> Variable

Tanh activation.

pub fn sum(&self) -> Variable

Sum all elements.

pub fn sum_dim(&self, dim: usize) -> Variable

Sum along a dimension, removing that dimension.

pub fn mean(&self) -> Variable

Mean of all elements.

pub fn mse_loss(&self, target: &Variable) -> Variable

Mean Squared Error loss.

pub fn binary_cross_entropy(&self, target: &Variable) -> Variable

Binary Cross Entropy loss (expects sigmoid output).

pub fn reshape(&self, shape: &[usize]) -> Variable

Reshapes the variable to a new shape.

pub fn transpose(&self, dim0: usize, dim1: usize) -> Variable

Transposes two dimensions.

pub fn slice(&self, ranges: &[Range<usize>]) -> Variable

Slices the variable along specified ranges.

pub fn narrow(&self, dim: usize, start: usize, length: usize) -> Variable

Narrows the variable along a dimension.

Returns a view of the tensor containing elements from start to start + length along the specified dimension. This operation preserves gradients for backpropagation.

pub fn expand(&self, shape: &[usize]) -> Variable

Expands the variable to a new shape (broadcast).

Tracks the computational graph for backward pass.

pub fn select(&self, dim: usize, index: usize) -> Variable

Selects a single index along a dimension, reducing rank by 1.

For a tensor of shape (A, B, C), select(1, i) returns shape (A, C). Tracks the computational graph for backward pass.

pub fn unsqueeze(&self, dim: usize) -> Variable

Adds a dimension of size 1 at the given position.

Tracks the computational graph for backward pass.

pub fn cat(variables: &[&Variable], dim: usize) -> Variable

Concatenates variables along a dimension.

All variables must have the same shape except along the cat dimension. Tracks the computational graph for backpropagation.

pub fn mul_scalar(&self, scalar: f32) -> Variable

Multiplies by a scalar.

pub fn add_scalar(&self, scalar: f32) -> Variable

Adds a scalar.

pub fn sub_scalar(&self, scalar: f32) -> Variable

Subtracts a scalar.

pub fn div_scalar(&self, scalar: f32) -> Variable

Divides by a scalar.

pub fn gelu(&self) -> Variable

GELU activation function (Gaussian Error Linear Unit).

pub fn silu(&self) -> Variable

SiLU/Swish activation function (x * sigmoid(x)).

pub fn sqrt(&self) -> Variable

Square root.

pub fn softmax(&self, dim: i32) -> Variable

Softmax along specified dimension.

pub fn log_softmax(&self, dim: i32) -> Variable

Log softmax along specified dimension.

pub fn mean_dim(&self, dim: i32, keepdim: bool) -> Variable

Mean along a dimension, optionally keeping the dimension.

pub fn var_dim(&self, dim: i32, keepdim: bool) -> Variable

Variance along a dimension, optionally keeping the dimension.

pub fn from_tensor_with_grad(data: Tensor<f32>, requires_grad: bool) -> Variable

Creates a Variable from a tensor and requires_grad flag (for weight access). This is typically used internally by Parameter types.

pub fn clone_var(&self) -> Variable

Clones the variable (alias for Clone trait).

pub fn add(&self, other: &Variable) -> Variable

Adds another variable (alias for add_var for method chaining).

pub fn sub(&self, other: &Variable) -> Variable

Subtracts another variable (alias for sub_var for method chaining).

pub fn mul(&self, other: &Variable) -> Variable

Multiplies by another variable (alias for mul_var for method chaining).

pub fn div(&self, other: &Variable) -> Variable

Divides by another variable (alias for div_var for method chaining).

Trait Implementations

impl Add for &Variable

type Output = Variable

The resulting type after applying the + operator.

fn add(self, other: &Variable) -> Variable

Performs the + operation.

impl Clone for Variable

fn clone(&self) -> Variable

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Variable

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.

impl Div for &Variable

type Output = Variable

The resulting type after applying the / operator.

fn div(self, other: &Variable) -> Variable

Performs the / operation.

impl Mul for &Variable

type Output = Variable

The resulting type after applying the * operator.

fn mul(self, other: &Variable) -> Variable

Performs the * operation.

impl Neg for &Variable

type Output = Variable

The resulting type after applying the - operator.

fn neg(self) -> Variable

Performs the unary - operation.

impl Sub for &Variable

type Output = Variable

The resulting type after applying the - operator.

fn sub(self, other: &Variable) -> Variable

Performs the - operation.