Features

• Flexible and intuitive custom neural network module 🤖
• Stateless and thread safe forward pass 🚀
• Fast training with full support for metric, logging and checkpoining 🌟
• Burn-Tensor: Tensor library with autodiff, CPU and GPU support 🔥
• Burn-Dataset: Dataset library with multiple utilities and sources 📚

Details

Example

Full example showing most of the features from burn MNIST.

Components

Knowing the main components will be of great help when starting playing with burn.

Backend

Almost everything is based on the Backend trait, which allows to run tensor operations with different implementations without having to change your code. A backend does not necessary have autodiff capabilities, therefore you can use ADBackend when you need it.

Tensor

The Tensor struct is at the core of the burn framework. It takes two generic parameters, the Backend and the number of dimensions D,

```rust use burn::tensor::{Tensor, Shape, Data}; use burn::tensor::backend::{NdArrayBackend, TchBackend};

let myndarraymatrix = Tensor::, 2>::ones(Shape::new([3, 3])); let mytchmatrix = Tensor::, 2>::from_data( Data::from([[1.0, 7.0], [13.0, -3.0]]) ); ```

Note that Data is not specific to any backend.

Module

The Module derive let your create your own neural network module similar to PyTorch.

```rust use burn::nn; use burn::module::{Param, Module}; use burn::tensor::backend::Backend;

[derive(Module, Debug)]

struct MyModule { my_param: Param>, repeat: usize, } ```

Note that only the fields wrapped inside Param are updated during training, and the other ones should implement Clone.

Forward

The Forward trait can also be implemented by your module.

```rust use burn::module::Forward; use burn::tensor::Tensor;

impl Forward, Tensor> for MyModule { fn forward(&self, input: Tensor) -> Tensor { let mut x = input;

   for _ in 0..self.repeat {
       x = self.my_param.forward(x);
   }

   x