Burn Application Development

Core knowledge for building applications with the Burn deep learning framework.

Tensors

Burn tensors are the fundamental data structure. Three element types:

• Tensor<B, D, Float> — Floating point operations
• Tensor<B, D, Int> — Integer operations
• Tensor<B, D, Bool> — Boolean masks

Key patterns:

// Creation
let tensor = Tensor::<B, 2>::zeros([batch, features], &device);
let tensor = Tensor::from_data([[1.0, 2.0], [3.0, 4.0]], &device);

// Operations (return new tensors, original unchanged)
let result = tensor.matmul(other);
let result = tensor.relu();

// Clone for multiple uses (cheap, reference counted)
let a = tensor.clone();
let b = tensor.clone();

Modules

Neural network layers use the Module derive macro:

#[derive(Module, Debug)]
pub struct Model<B: Backend> {
    conv: Conv2d<B>,
    pool: AdaptiveAvgPool2d,
    linear: Linear<B>,
    activation: Relu,
}

impl<B: Backend> Model<B> {
    pub fn forward(&self, x: Tensor<B, 4>) -> Tensor<B, 2> {
        let x = self.conv.forward(x);
        let x = self.activation.forward(x);
        let x = self.pool.forward(x);
        let x = x.flatten(1, 3);
        self.linear.forward(x)
    }
}

Config

Type-safe configuration with the Config derive:

#[derive(Config)]
pub struct ModelConfig {
    #[config(default = 64)]
    hidden_size: usize,
    #[config(default = 0.1)]
    dropout: f64,
}

// Usage
let config = ModelConfig::new();
let model = config.init::<B>(&device);

Autodiff

Automatic differentiation for training:

// AutodiffBackend wraps any backend
type MyBackend = Autodiff<Wgpu>;

// Forward pass tracks gradients
let output = model.forward(input);
let loss = output.cross_entropy(targets);

// Backward pass
let grads = loss.backward();
let grad_tensor = tensor.grad(&grads).unwrap();

Key difference from PyTorch: gradients are returned as a separate Gradients struct, not stored on tensors.

Records

Serialization for model weights:

// Save
let recorder = CompactRecorder::new();
model.save_file("model.bin", &recorder)?;

// Load
let model = config.init::<B>(&device);
let model = model.load_file("model.bin", &recorder, &device)?;

Additional Resources

Consult references/topic-map-app.md for:

• Detailed tensor operation reference
• Built-in module catalog (Conv, Pool, RNN, Transformer, Loss)
• Advanced autodiff patterns
• Record format options