Struct Tensor 

pub struct Tensor {
    pub data: Array,
    pub grad: Option<Rc<RefCell<Array>>>,
    pub requires_grad: bool,
    pub compute_node: Option<Rc<ComputeNode>>,
    pub is_leaf: bool,
}

Tensor con soporte para autograd

Fields

data: Array

Datos del tensor

grad: Option<Rc<RefCell<Array>>>

Gradiente acumulado (dL/d(este tensor))

requires_grad: bool

Si este tensor requiere gradientes

compute_node: Option<Rc<ComputeNode>>

Nodo del compute graph (None para leaf tensors)

is_leaf: bool

Si es un leaf tensor (parámetro del modelo)

Implementations

impl Tensor

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

Add: self + other

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

Sub: self - other

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

Mul (elementwise): self * other

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

Div: self / other

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

MatMul: self @ other

pub fn relu(&self) -> Result<Self>

ReLU: max(0, self)

pub fn sigmoid(&self) -> Result<Self>

Sigmoid: 1 / (1 + exp(-self))

pub fn exp(&self) -> Result<Self>

Exp: exp(self)

pub fn log(&self) -> Result<Self>

Log: log(self)

pub fn sum(&self) -> Result<Self>

Sum: suma todos los elementos

pub fn mean(&self) -> Result<Self>

Mean: promedio de todos los elementos

pub fn mse_loss(&self, target: &Tensor) -> Result<Self>

MSE Loss: mean((self - target)^2)

pub fn cross_entropy_loss(&self, target: &Tensor) -> Result<Self>

Cross-entropy loss con softmax integrado (Batch-aware)

pub fn flatten(&self, start_dim: usize, end_dim: usize) -> Result<Self>

Flatten: flatten(start, end)

pub fn reshape(&self, shape: Vec<usize>) -> Result<Self>

Reshape: input.reshape([d1, d2, …])

pub fn conv1d( &self, weight: &Tensor, bias: Option<&Tensor>, stride: usize, padding: usize, ) -> Result<Self>

Conv1D

pub fn batch_norm( &self, running_mean: &mut Tensor, running_var: &mut Tensor, weight: &Tensor, bias: &Tensor, training: bool, momentum: f32, eps: f32, ) -> Result<Self>

BatchNorm

pub fn dropout(&self, p: f32, training: bool) -> Result<Self>

Dropout

pub fn pow(&self, exponent: f32) -> Result<Self>

Pow: self^exponent

pub fn sqrt(&self) -> Result<Self>

Sqrt: sqrt(self)

pub fn sin(&self) -> Result<Self>

Sin: sin(self)

pub fn cos(&self) -> Result<Self>

Cos: cos(self)

pub fn tan(&self) -> Result<Self>

Tan: tan(self)

pub fn tanh(&self) -> Result<Self>

Tanh: tanh(self)

impl Tensor

pub fn new(data: Array, requires_grad: bool) -> Self

Crea un nuevo tensor desde un Array

Argumentos

• data: Array con los datos
• requires_grad: Si se deben calcular gradientes para este tensor

pub fn from_operation( data: Array, compute_node: ComputeNode, requires_grad: bool, ) -> Self

Crea tensor desde resultado de operación

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

Acceso directo a la forma

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

Acceso directo a los datos

pub fn gradient(&self) -> Option<Array>

Obtiene el gradiente actual (si existe)

pub fn item(&self) -> f32

Obtiene el valor escalar de un tensor de 1 elemento

pub fn set_gradient(&self, grad: Array) -> Result<()>

Establece el gradiente manualmente

pub fn accumulate_gradient(&self, incoming_grad: &Array) -> Result<()>

Acumula gradiente (grad += incoming_grad)

pub fn zero_grad(&self)

Resetea el gradiente a cero

pub fn detach(&self) -> Self

Desconecta del compute graph (detach) Retorna un nuevo tensor sin gradientes

pub fn to_array(&self) -> Array

Convierte a Array (sin autograd)

pub fn node_id(&self) -> Option<NodeId>

Obtiene el ID del nodo en el compute graph

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

Backpropagation - Calcula gradientes desde este tensor hacia las hojas

Este es el algoritmo principal de autograd:

1. Inicializa grad = 1.0 para este tensor (dL/dL = 1)
2. Recorre el graph en orden topológico inverso
3. Para cada operación, llama su backward_fn
4. Acumula gradientes en los inputs

pub fn transpose(&self) -> Result<Tensor>

Transpose del tensor (invierte dimensiones)

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.

impl Display for Tensor

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

Formats the value using the given formatter.