Burn Optimizers

Optimizers and learning rate schedulers for Burn

Applications use these through burn::optim, burn::lr_scheduler and burn::grad_clipping. An optimizer is built from its config and applied to a module with gradients from a backward pass:

use burn::optim::{AdamConfig, GradientsParams};

let mut optimizer = AdamConfig::new().init();
let grads = GradientsParams::from_grads(loss.backward(), &model);
model = optimizer.step(learning_rate, model, grads);

• Optimizers: SGD, Adam, AdamW, Adagrad, Adafactor, Adan, LAMB, L-BFGS, Lion, Muon and RMSprop, with momentum, weight decay and gradient accumulation helpers.
• lr_scheduler: constant, step, exponential, linear, cosine, Noam, and sequential or composed schedules.
• grad_clipping: clipping by value or by norm.

See the optimizer and learning rate scheduler chapters 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.