Burn Neural Networks

Neural network layers, activations and losses for Burn

Every layer is a module built from a config, and applications use them through burn::nn:

use burn::nn::{Linear, LinearConfig};

let linear: Linear = LinearConfig::new(784, 128).init(&device);
let output = linear.forward(input);

• Layers: linear, convolution and transposed convolution (1D to 3D), pooling, normalization (batch, layer, group, instance, RMS), embeddings, dropout, recurrent layers (LSTM, GRU), attention and transformers, positional and rotary encodings, interpolation, and more.
• activation: activation functions as modules.
• loss: loss functions, from mean squared error and cross-entropy to CTC.
• Initializer: weight initialization schemes.

See the module chapter of the Burn Book.

Feature Flags

• std (default): standard library support. Without it the crate is no_std with alloc.
• tracing: instrument operations with the tracing crate.

Part of the Burn deep learning framework. See the Burn Book and the API documentation.