Struct Tensor 

pub struct Tensor { /* private fields */ }

A tensor with optional gradient tracking for automatic differentiation.

Design

The tensor stores:

• data: The actual numerical values (backed by aprender’s Vector)
• shape: Dimensions of the tensor
• grad: Accumulated gradient (populated after backward())
• requires_grad: Whether this tensor participates in gradient computation
• grad_fn: The operation that created this tensor (for backprop)
• id: Unique identifier for graph tracking

Thread Safety

Tensors use Arc internally for shared ownership of gradient functions, making them safe to share across threads for inference (but not training).

Implementations

impl Tensor

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

Element-wise addition: z = self + other

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

Element-wise subtraction: z = self - other

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

Element-wise multiplication: z = self * other

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

Element-wise division: z = self / other

pub fn neg(&self) -> Tensor

Element-wise negation: z = -self

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

Scalar multiplication: z = self * scalar

impl Tensor

pub fn exp(&self) -> Tensor

Element-wise exponential: z = exp(self)

pub fn log(&self) -> Tensor

Element-wise natural logarithm: z = log(self)

pub fn pow(&self, n: f32) -> Tensor

Element-wise power: z = self^n

pub fn sqrt(&self) -> Tensor

Element-wise square root: z = sqrt(self)

impl Tensor

pub fn sum(&self) -> Tensor

Sum all elements: z = sum(self)

pub fn mean(&self) -> Tensor

Mean of all elements: z = mean(self)

impl Tensor

pub fn relu(&self) -> Tensor

ReLU activation: z = max(0, self)

pub fn sigmoid(&self) -> Tensor

Sigmoid activation: z = 1 / (1 + exp(-self))

pub fn tanh_(&self) -> Tensor

Tanh activation

pub fn leaky_relu(&self, negative_slope: f32) -> Tensor

Leaky ReLU activation: z = max(negative_slope * x, x)

Arguments

• negative_slope - Controls the angle of the negative slope (default: 0.01)

pub fn gelu(&self) -> Tensor

GELU (Gaussian Error Linear Unit) activation.

Uses the tanh approximation: GELU(x) ≈ 0.5 * x * (1 + tanh(sqrt(2/π) * (x + 0.044715 * x³)))

pub fn softmax(&self) -> Tensor

Softmax activation over the last dimension of a 2D tensor.

softmax(x)_i = exp(x_i) / Σ_j exp(x_j)

Uses numerically stable computation with max subtraction.

impl Tensor

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

Matrix multiplication: z = self @ other

Currently supports 2D tensors only. Batched matmul (3D+ tensors) can be added by iterating over batch dimensions and calling 2D matmul.

pub fn transpose(&self) -> Tensor

Transpose a 2D tensor.

Example

let a = Tensor::new(&[1.0, 2.0, 3.0, 4.0], &[2, 2]);
let a_t = a.transpose();
// a_t = [[1, 3], [2, 4]]

pub fn broadcast_add(&self, other: &Tensor) -> Tensor

Broadcast addition: z = matrix + vector (broadcasts over rows).

The vector is broadcast to match the matrix’s second dimension. This is useful for adding biases in neural networks.

Shape

• self: [N, M] (2D matrix)
• other: [M] (1D vector)
• output: [N, M]

Example

let matrix = Tensor::new(&[1.0, 2.0, 3.0, 4.0], &[2, 2]);
let bias = Tensor::new(&[10.0, 20.0], &[2]);
let result = matrix.broadcast_add(&bias);
// result = [[11, 22], [13, 24]]

pub fn view(&self, new_shape: &[usize]) -> Tensor

Reshape tensor to a new shape (view).

The total number of elements must remain the same.

Example

let a = Tensor::new(&[1.0, 2.0, 3.0, 4.0, 5.0, 6.0], &[2, 3]);
let b = a.view(&[3, 2]);
// b = [[1, 2], [3, 4], [5, 6]]

impl Tensor

pub fn new(data: &[f32], shape: &[usize]) -> Self

Create a new tensor from a slice with the given shape.

By default, gradient tracking is disabled.

Panics

Panics if the data length doesn’t match the product of shape dimensions.

pub fn from_slice(data: &[f32]) -> Self

Create a tensor from a 1D slice (vector).

pub fn zeros(shape: &[usize]) -> Self

Create a tensor filled with zeros.

pub fn ones(shape: &[usize]) -> Self

Create a tensor filled with ones.

pub fn zeros_like(other: &Tensor) -> Self

Create a tensor with the same shape as another, filled with zeros.

pub fn ones_like(other: &Tensor) -> Self

Create a tensor with the same shape as another, filled with ones.

pub fn requires_grad(self) -> Self

Enable gradient tracking for this tensor.

Returns self for method chaining.

pub fn requires_grad_(&mut self, requires: bool) -> &mut Self

Enable or disable gradient tracking (in-place).

pub fn requires_grad_enabled(&self) -> bool

Check if this tensor requires gradient computation.

pub fn is_leaf(&self) -> bool

Check if this is a leaf tensor (not created by an operation).

pub fn id(&self) -> TensorId

Get the tensor’s unique identifier.

pub fn shape(&self) -> &[usize]

Get the shape of the tensor.

pub fn numel(&self) -> usize

Get the total number of elements.

pub fn ndim(&self) -> usize

Get the number of dimensions.

pub fn data(&self) -> &[f32]

Get a reference to the underlying data.

pub fn data_mut(&mut self) -> &mut [f32]

Get a mutable reference to the underlying data.

Warning

Modifying data directly may invalidate gradients.

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

Get the gradient tensor (if computed).

pub fn zero_grad_(&mut self)

Zero out the gradient.

pub fn clear_grad(&mut self)

Clear the gradient (alias for zero_grad_).

pub fn detach(&self) -> Tensor

Detach tensor from computation graph.

Returns a new tensor with the same data but no gradient tracking.

pub fn item(&self) -> f32

Get a scalar value (for 0-d or 1-element tensors).

Panics

Panics if the tensor has more than one element.

pub fn backward(&self)

Compute gradients via backpropagation.

This implements the reverse-mode automatic differentiation algorithm described in Rumelhart et al. (1986).

Panics

Panics if called on a tensor with more than one element (use backward_with_grad for non-scalar outputs).

pub fn backward_with_grad(&self, grad_output: Tensor)

Compute gradients with a specified output gradient.

Arguments

• grad_output - Gradient of the loss with respect to this tensor

Trait Implementations

impl Clone for Tensor

fn clone(&self) -> Tensor

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Tensor

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

Formats the value using the given formatter.