Struct burn_tensor::Tensor

pub struct Tensor<B, const D: usize, K = Float>where
    B: Backend,
    K: TensorKind<B>,{ /* private fields */ }

Implementations

impl<B, const D: usize, K> Tensor<B, D, K>where B: Backend, K: TensorKind<B>,

pub fn new(primitive: K::Primitive<D>) -> Self

Constructs a new Tensor.

impl<B, const D: usize, K> Tensor<B, D, K>where B: Backend, K: BasicOps<B>,

pub fn empty<S: Into<Shape<D>>>(shape: S) -> Self

Create an empty tensor of the given shape.

pub fn empty_device<S: Into<Shape<D>>>(shape: S, device: &B::Device) -> Self

Create an empty tensor of the given shape.

pub fn dims(&self) -> [usize; D]

Returns the dimensions of the current tensor.

Equivalent to tensor.shape().dims.

pub fn shape(&self) -> Shape<D>

Returns the shape of the current tensor.

pub fn reshape<const D2: usize, S: Into<Shape<D2>>>( self, shape: S ) -> Tensor<B, D2, K>

Reshape the tensor to have the given shape.

Panics

If the tensor can not be reshape to the given shape.

pub fn flatten<const D2: usize>( self, start_dim: usize, end_dim: usize ) -> Tensor<B, D2, K>

Flatten the tensor along a given range of dimensions.

This function collapses the specified range of dimensions into a single dimension, effectively flattening the tensor in that range.

Arguments

• start_dim: The starting dimension of the range to be flattened.
• end_dim: The ending dimension of the range to be flattened (inclusive).

Type Parameters

• D2: The resulting number of dimensions in the flattened tensor.

Returns

A new Tensor<B, D2, K> instance with the specified range of dimensions flattened.

Example

use burn_tensor::backend::Backend;
use burn_tensor::{Tensor, Shape};

fn example<B: Backend>() {
    let tensor = Tensor::<B, 3>::ones(Shape::new([2, 3, 4]));

    // Given a 3D tensor with dimensions (2, 3, 4), flatten the dimensions between indices 1 and 2:
    let flattened_tensor: Tensor::<B, 2> = tensor.flatten(1, 2);

    // The resulting tensor will have dimensions (2, 12).
   println!("{:?}", flattened_tensor.shape());
}

pub fn unsqueeze<const D2: usize>(self) -> Tensor<B, D2, K>

Unsqueeze the current tensor. Create new dimensions to fit the given size.

Panics

If the output size is higher than the current tensor.

Example

use burn_tensor::backend::Backend;
use burn_tensor::{Tensor, Shape};

fn example<B: Backend>() {
    let tensor = Tensor::<B, 2>::ones(Shape::new([3, 3]));
    let tensor = tensor.unsqueeze::<4>();
    println!("{:?}", tensor.shape());
    // Shape { dims: [1, 1, 3, 3] }
}

pub fn index<const D2: usize>(self, indexes: [Range<usize>; D2]) -> Self

Returns a tensor containing the elements selected from the given ranges.

Panics

If a range exceeds the number of elements on a dimension.

Example

use burn_tensor::backend::Backend;
use burn_tensor::{Tensor, Shape};

fn example<B: Backend>() {
    let tensor = Tensor::<B, 3>::ones(Shape::new([2, 3, 3]));
    let tensor_indexed = tensor.index([0..1, 0..3, 1..2]);
    println!("{:?}", tensor_indexed.shape());
    // Shape { dims: [1, 3, 2] }
}

pub fn index_assign<const D2: usize>( self, indexes: [Range<usize>; D2], values: Self ) -> Self

Returns a copy of the current tensor with the selected elements changed to the new ones at the selected indexes.

Panics

• If a range exceeds the number of elements on a dimension.
• If the given values don’t match the given ranges.

Example

use burn_tensor::backend::Backend;
use burn_tensor::Tensor;

fn example<B: Backend>() {
    let tensor = Tensor::<B, 3>::ones([2, 3, 3]);
    let values = Tensor::<B, 3>::zeros([1, 1, 1]);
    let tensor_indexed = tensor.index_assign([0..1, 0..1, 0..1], values);
    println!("{:?}", tensor_indexed.shape());
    // Shape { dims: [2, 3, 3] }
}

pub fn device(&self) -> B::Device

Returns the device of the current tensor.

pub fn to_device(self, device: &B::Device) -> Self

Returns a new tensor on the given device.

pub fn into_data(self) -> Data<K::Elem, D>

Returns the data of the current tensor.

pub fn to_data(&self) -> Data<K::Elem, D>

Returns the data of the current tensor without taking ownership.

pub fn from_data<T>(data: T) -> Selfwhere T: Into<Data<K::Elem, D>>,

Create a tensor from the given data.

pub fn from_data_device<T>(data: T, device: &B::Device) -> Selfwhere T: Into<Data<K::Elem, D>>,

Create a tensor from the given data on the given device.

pub fn repeat(self, dim: usize, times: usize) -> Self

Repeat the tensor along the given dimension.

Panics

If the selected dimension more than one item.

pub fn equal(self, other: Self) -> Tensor<B, D, Bool>

Applies element wise equal comparison and returns a boolean tensor.

Panics

If the two tensors don’t have the same shape.

pub fn equal_elem<E: Into<K::Elem>>(self, other: E) -> Tensor<B, D, Bool>

Applies element wise equal comparison and returns a boolean tensor.

pub fn cat(tensors: Vec<Self>, dim: usize) -> Self

Concatenates all tensors into a new one along the given dimension.

Panics

If all tensors don’t have the same shape.

impl<B, const D: usize> Tensor<B, D, Bool>where B: Backend,

pub fn from_bool(data: Data<bool, D>) -> Self

Create a boolean tensor from data.

pub fn from_bool_device(data: Data<bool, D>, device: &B::Device) -> Self

Create a boolean tensor from data on the given device.

pub fn into_int(self) -> Tensor<B, D, Int>

Convert the bool tensor into an int tensor.

impl<const D: usize, B> Tensor<B, D>where B: Backend,

pub fn into_primitive(self) -> B::TensorPrimitive<D>

pub fn from_primitive(tensor: B::TensorPrimitive<D>) -> Self

pub fn inplace<F: FnOnce(Self) -> Self>(&mut self, func: F)

Executes an operation on the tensor and modifies its value.

Notes

This won’t necessary reuse the same tensor data/buffer, but it should if there is no other reference pointing to the same tensor.

Wrapping operations with inplace is not an optimization, it’s mainly there if you want to mutate a tensor by using owned operations. A plausible usage would be to update the weights of a mutable model reference.

pub fn exp(self) -> Self

Applies element wise exponential operation.

y = e^x

pub fn log(self) -> Self

Applies element wise natural log operation ln.

y = log(x)

pub fn log1p(self) -> Self

Applies the natural logarithm of one plus the input tensor, element-wise.

y = log(x+1)

pub fn erf(self) -> Self

Applies the error function element wise.

y = erf(x)

pub fn powf(self, value: f32) -> Self

Applies element wise power operation.

y = x^a

pub fn sqrt(self) -> Self

Applies element wise root square operation.

pub fn cos(self) -> Self

Applies element wise cosine operation.

pub fn sin(self) -> Self

Applies element wise sine operation.

pub fn tanh(self) -> Self

Applies element wise hyperbolic tangent operation.

pub fn from_floats<A: Into<Data<f32, D>>>(floats: A) -> Self

Create a tensor from floats (f32).

Example

use burn_tensor::backend::Backend;
use burn_tensor::Tensor;

fn example<B: Backend>() {
    let _ = Tensor::<B, 1>::from_floats([1.0, 2.0]);
    let _ = Tensor::<B, 2>::from_floats([[1.0, 2.0], [3.0, 4.0]]);
}

pub fn zeros_like(&self) -> Self

Returns a new tensor with the same shape and device as the current tensor filled with zeros.

pub fn ones_like(&self) -> Self

Returns a new tensor with the same shape and device as the current tensor filled with ones.

pub fn random_like(&self, distribution: Distribution<B::FloatElem>) -> Self

Returns a new tensor with the same shape and device as the current tensor filled random values sampled from the given distribution.

pub fn one_hot(index: usize, num_classes: usize) -> Self

Create a one hot tensor.

Example

use burn_tensor::backend::Backend;
use burn_tensor::Tensor;

fn example<B: Backend>() {
    let one_hot = Tensor::<B, 1>::one_hot(2, 10);
    println!("{}", one_hot.to_data());
    // [0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
}

pub fn transpose(self) -> Self

Applies the transpose operation.

On matrix and higher dimension tensor, it swap the last two dimensions.

Panics

If the tensor is of 1 dimension or less.

pub fn swap_dims(self, dim1: usize, dim2: usize) -> Self

Swap two dimensions.

Panics

If the dimensions exceed the shape of than the tensor.

pub fn matmul(self, other: Self) -> Self

Applies the matrix multiplication operation.

C = AB

Panics

If the two tensors dont’ have a compatible shape.

pub fn var(self, dim: usize) -> Self

Calculate the variance along the given dimension.

pub fn var_bias(self, dim: usize) -> Self

Calculate the variance along the given dimension without applying the Bessel’s correction.

pub fn var_mean(self, dim: usize) -> (Self, Self)

Calculate the variance along the given dimension and also returns the mean.

pub fn var_mean_bias(self, dim: usize) -> (Self, Self)

Calculate the variance along the given dimension without applying the Bessel’s correction and also returns the mean.

pub fn random<S: Into<Shape<D>>>( shape: S, distribution: Distribution<B::FloatElem> ) -> Self

Create a random tensor of the given shape where each element is sampled from the given distribution.

pub fn to_full_precision(&self) -> Tensor<B::FullPrecisionBackend, D>

Returns a tensor with full precision based on the selected backend.

pub fn from_full_precision(tensor: Tensor<B::FullPrecisionBackend, D>) -> Self

Returns a tensor on the selected backend from a full precision tensor.

pub fn argmax(self, dim: usize) -> Tensor<B, D, Int>

Applies the argmax function along the given dimension and returns an integer tensor.

Example

use burn_tensor::backend::Backend;
use burn_tensor::{Tensor, Shape};

fn example<B: Backend>() {
    let tensor = Tensor::<B, 3>::ones(Shape::new([2, 3, 3]));
    let tensor = tensor.argmax(1);
    println!("{:?}", tensor.shape());
    // Shape { dims: [2, 1, 3] }
}

pub fn argmin(self, dim: usize) -> Tensor<B, D, Int>

Applies the argmin function along the given dimension and returns an integer tensor.

Example

use burn_tensor::backend::Backend;
use burn_tensor::{Tensor, Shape};

fn example<B: Backend>() {
    let tensor = Tensor::<B, 3>::ones(Shape::new([2, 3, 3]));
    let tensor = tensor.argmin(1);
    println!("{:?}", tensor.shape());
    // Shape { dims: [2, 1, 3] }
}

pub fn detach(self) -> Self

Detach the current tensor from the autodiff graph. This function does nothing when autodiff is not enabled. This can be used in batchers or elsewere to ensure that previous operations are not considered in the autodiff graph.

pub fn require_grad(self) -> Self

Mark the tensor to keep gradients during the backward pass. This function does nothing when autodiff is not enabled.

pub fn is_require_grad(&self) -> bool

Returns true if the tensor requires gradients during the backward pass.

pub fn set_require_grad(self, require_grad: bool) -> Self

Mark the tensor as tracked or untracked depending on the require grad argument. When tracked, the gradients will be available after the backward pass.

This function does nothing when autodiff is not enabled.

impl<const D: usize, B: ADBackend> Tensor<B, D>

pub fn backward(&self) -> B::Gradients

pub fn grad(&self, grads: &B::Gradients) -> Option<Tensor<B::InnerBackend, D>>

Get the gradients of a tensor if it exist.

Returns a new reference to the same tensor. Therefore the same grad tensor can be accessed multiple times. If you only need to get the gradients one time, consider using grad_remove for better performance.

pub fn grad_remove( &self, grads: &mut B::Gradients ) -> Option<Tensor<B::InnerBackend, D>>

Remove the grad tensor from the grads struct returning the result.

pub fn inner(self) -> Tensor<B::InnerBackend, D>

pub fn from_inner(inner: Tensor<B::InnerBackend, D>) -> Self

impl<B> Tensor<B, 1, Int>where B: Backend,

pub fn arange(range: Range<usize>) -> Self

Returns a new integer tensor on the default device which values are generated from the given range.

pub fn arange_device(range: Range<usize>, device: &B::Device) -> Self

Returns a new integer tensor on the specified device which values are generated from the given range.

impl<B, const D: usize, K> Tensor<B, D, K>where B: Backend, K: Numeric<B>, K::Elem: Element,

pub fn into_scalar(self) -> K::Elem

Convert the tensor into a scalar.

Panics

If the tensor doesn’t have one element.

pub fn add(self, other: Self) -> Self

Applies element wise addition operation.

y = x2 + x1

pub fn add_scalar<E: ElementConversion>(self, other: E) -> Self

Applies element wise addition operation with a scalar.

y = x + s

pub fn sub(self, other: Self) -> Self

Applies element wise substraction operation.

y = x2 - x1

pub fn sub_scalar<E: ElementConversion>(self, other: E) -> Self

Applies element wise substraction operation with a scalar.

y = x - s

pub fn div(self, other: Self) -> Self

Applies element wise division operation.

y = x2 / x1

pub fn div_scalar<E: ElementConversion>(self, other: E) -> Self

Applies element wise division operation with a scalar.

y = x / s

pub fn mul(self, other: Self) -> Self

Applies element wise multiplication operation.

y = x2 * x1

pub fn mul_scalar<E: ElementConversion>(self, other: E) -> Self

Applies element wise multiplication operation with a scalar.

y = x * s

pub fn neg(self) -> Self

Switch sign of each element in the tensor.

y = -x

pub fn zeros<S: Into<Shape<D>>>(shape: S) -> Self

Create a tensor of the given shape where each element is zero.

pub fn zeros_device<S: Into<Shape<D>>>(shape: S, device: &B::Device) -> Self

Create a tensor of the given shape where each element is zero.

pub fn ones<S: Into<Shape<D>>>(shape: S) -> Self

Create a tensor of the given shape where each element is one.

pub fn ones_device<S: Into<Shape<D>>>(shape: S, device: &B::Device) -> Self

Create a tensor of the given shape where each element is zero.

pub fn mean(self) -> Tensor<B, 1, K>

Aggregate all elements in the tensor with the mean operation.

pub fn sum(self) -> Tensor<B, 1, K>

Aggregate all elements in the tensor with the sum operation.

pub fn mean_dim(self, dim: usize) -> Self

Aggregate all elements along the given dimension or axis in the tensor with the mean operation.

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

Aggregate all elements along the given dimension or axis in the tensor with the sum operation.

pub fn greater(self, other: Self) -> Tensor<B, D, Bool>

Applies element wise greater comparison and returns a boolean tensor.

Panics

If the two tensors don’t have the same shape.

pub fn greater_equal(self, other: Self) -> Tensor<B, D, Bool>

Applies element wise greater-equal comparison and returns a boolean tensor.

Panics

If the two tensors don’t have the same shape.

pub fn lower(self, other: Self) -> Tensor<B, D, Bool>

Applies element wise lower comparison and returns a boolean tensor.

Panics

If the two tensors don’t have the same shape.

pub fn lower_equal(self, other: Self) -> Tensor<B, D, Bool>

Applies element wise lower-equal comparison and returns a boolean tensor.

Panics

If the two tensors don’t have the same shape.

pub fn greater_elem<E: ElementConversion>(self, other: E) -> Tensor<B, D, Bool>

Applies element wise greater comparison and returns a boolean tensor.

pub fn greater_equal_elem<E: ElementConversion>( self, other: E ) -> Tensor<B, D, Bool>

Applies element wise greater-equal comparison and returns a boolean tensor.

pub fn lower_elem<E: ElementConversion>(self, other: E) -> Tensor<B, D, Bool>

Applies element wise lower comparison and returns a boolean tensor.

pub fn lower_equal_elem<E: ElementConversion>( self, other: E ) -> Tensor<B, D, Bool>

Applies element wise lower-equal comparison and returns a boolean tensor.

pub fn mask_scatter(self, mask: Tensor<B, D, Bool>, source: Self) -> Self

Fill elements from the given tensor based where the mask is true.

pub fn mask_fill<E: ElementConversion>( self, mask: Tensor<B, D, Bool>, value: E ) -> Self

Fill each element with the given value based on the given mask.

pub fn index_select(self, indexes: Tensor<B, D, Int>) -> Self

Select the tensor elements corresponding to the given indexes.

Notes

The index tensor shoud have the same shape as the original tensor except for the last dimension.

pub fn index_select_assign( self, indexes: Tensor<B, D, Int>, values: Self ) -> Self

Assign the selected elements corresponding to the given indexes from the value tensor to the original tensor using sum reduction.

Notes

The index tensor shoud have the same shape as the original tensor except for the last dimension. The value and index tensors should have the same shape.

pub fn index_select_dim(self, dim: usize, indexes: Tensor<B, 1, Int>) -> Self

Select the tensor elements along the given dimension corresponding to the given indexes.

pub fn index_select_dim_assign<const D2: usize>( self, dim: usize, indexes: Tensor<B, 1, Int>, values: Tensor<B, D2, K> ) -> Self

Assign the selected elements along the given dimension corresponding to the given indexes from the value tensor to the original tensor using sum reduction.

Trait Implementations

impl<E, const D: usize, B, K> Add<E> for Tensor<B, D, K>where E: ElementConversion, B: Backend, K: Numeric<B>, K::Elem: Element,

type Output = Tensor<B, D, K>

The resulting type after applying the + operator.

fn add(self, other: E) -> Self

Performs the + operation.

impl<B, const D: usize, K> Add<Tensor<B, D, K>> for Tensor<B, D, K>where B: Backend, K: Numeric<B>, K::Elem: Element,

type Output = Tensor<B, D, K>

The resulting type after applying the + operator.

fn add(self, rhs: Tensor<B, D, K>) -> Self

Performs the + operation.

impl<B, const D: usize, K> Clone for Tensor<B, D, K>where B: Backend + Clone, K: TensorKind<B> + Clone, K::Primitive<D>: Clone,

fn clone(&self) -> Tensor<B, D, K>

Returns a copy of the value.

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

Performs copy-assignment from source.

impl<B, const D: usize, K> Debug for Tensor<B, D, K>where B: Backend + Debug, K: TensorKind<B> + Debug, K::Primitive<D>: Debug,

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

Formats the value using the given formatter.

impl<B, const D: usize, K> Display for Tensor<B, D, K>where B: Backend, B::IntElem: Display, K: BasicOps<B>, <K as BasicOps<B>>::Elem: Debug,

Pretty print tensors

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

Formats the value using the given formatter.

impl<E, const D: usize, B, K> Div<E> for Tensor<B, D, K>where E: ElementConversion, B: Backend, K: Numeric<B>, K::Elem: Element,

type Output = Tensor<B, D, K>

The resulting type after applying the / operator.

fn div(self, other: E) -> Self

Performs the / operation.

impl<B, const D: usize, K> Div<Tensor<B, D, K>> for Tensor<B, D, K>where B: Backend, K: Numeric<B>, K::Elem: Element,

type Output = Tensor<B, D, K>

The resulting type after applying the / operator.

fn div(self, rhs: Tensor<B, D, K>) -> Self

Performs the / operation.

impl<E, const D: usize, B, K> Mul<E> for Tensor<B, D, K>where E: ElementConversion, B: Backend, K: Numeric<B>, K::Elem: Element,

type Output = Tensor<B, D, K>

The resulting type after applying the * operator.

fn mul(self, other: E) -> Self

Performs the * operation.

impl<B, const D: usize, K> Mul<Tensor<B, D, K>> for Tensor<B, D, K>where B: Backend, K: Numeric<B>, K::Elem: Element,

type Output = Tensor<B, D, K>

The resulting type after applying the * operator.

fn mul(self, rhs: Tensor<B, D, K>) -> Self

Performs the * operation.

impl<B, const D: usize, K> Neg for Tensor<B, D, K>where B: Backend, K: Numeric<B>, K::Elem: Element,

type Output = Tensor<B, D, K>

The resulting type after applying the - operator.

fn neg(self) -> Self

Performs the unary - operation.

impl<E, const D: usize, B, K> Sub<E> for Tensor<B, D, K>where E: ElementConversion, B: Backend, K: Numeric<B>, K::Elem: Element,

type Output = Tensor<B, D, K>

The resulting type after applying the - operator.

fn sub(self, other: E) -> Self

Performs the - operation.

impl<B, const D: usize, K> Sub<Tensor<B, D, K>> for Tensor<B, D, K>where B: Backend, K: Numeric<B>, K::Elem: Element,

type Output = Tensor<B, D, K>

The resulting type after applying the - operator.

fn sub(self, rhs: Tensor<B, D, K>) -> Self

Performs the - operation.