Implementation:LaurentMazare Tch rs Adam Build

Knowledge Sources	tch-rs
Domains	Deep_Learning, Optimization
Last Updated	2026-02-08 14:00 GMT

Overview

Concrete tool for constructing the Adam optimizer for gradient-based training provided by the tch nn module.

Description

nn::Adam::default().build(&vs, lr) creates an Optimizer wrapping a C++ Adam optimizer with all trainable variables from the given VarStore registered for optimization. The optimizer provides backward_step for combined zero_grad + backward + step operations, and set_lr for learning rate scheduling.

Usage

Use after creating a VarStore and defining all model layers. Pass the VarStore reference and a learning rate to bind the optimizer to all trainable parameters.

Code Reference

Source Location

• Repository: tch-rs
• File: src/nn/optimizer.rs
• Lines: 73-76 (Adam::default), 23-34 (OptimizerConfig::build)

Signature

// Adam configuration
impl Default for Adam {
    fn default() -> Self {
        Adam { beta1: 0.9, beta2: 0.999, wd: 0., eps: 1e-8, amsgrad: false }
    }
}

// OptimizerConfig::build (shared by all optimizers)
fn build(self, vs: &VarStore, lr: f64) -> Result<Optimizer, TchError>

Import

use tch::nn::{self, OptimizerConfig};

I/O Contract

Inputs

Name	Type	Required	Description
vs	&VarStore	Yes	Variable store whose trainable parameters to optimize
lr	f64	Yes	Learning rate

Outputs

Name	Type	Description
Result<Optimizer>	nn::Optimizer	Optimizer wrapping C++ Adam with all trainable variables registered

Usage Examples

use tch::nn::{self, OptimizerConfig};

let vs = nn::VarStore::new(tch::Device::Cpu);
// ... define model layers using vs ...
let mut opt = nn::Adam::default().build(&vs, 1e-3)?;

// Training step
let loss = model.forward(&input).cross_entropy_for_logits(&labels);
opt.backward_step(&loss);

Related Pages

Implements Principle

• Principle:LaurentMazare_Tch_rs_Adam_Optimization