Struct AdversarialTrainer 

pub struct AdversarialTrainer { /* private fields */ }

Adversarial training wrapper for dual-network (generator/discriminator) setups.

Manages alternating training of two networks with configurable step ratios.

Example

use axonml::adversarial::AdversarialTrainer;

let mut trainer = AdversarialTrainer::new(gen_optimizer, disc_optimizer);
trainer.set_disc_steps_per_gen(5); // Train D 5x per G step

for batch in data {
    let (g_loss, d_loss) = trainer.step(&gen, &disc, &real, &noise, |pred, target| {
        bce_loss(pred, target)
    });
}

Implementations

impl AdversarialTrainer

pub fn new( gen_optimizer: Box<dyn Optimizer>, disc_optimizer: Box<dyn Optimizer>, ) -> Self

Creates a new adversarial trainer.

pub fn set_disc_steps_per_gen(&mut self, steps: usize)

Sets how many discriminator steps per generator step.

pub fn step<G, D, F>( &mut self, generator: &G, discriminator: &D, real_data: &Variable, noise: &Variable, loss_fn: F, ) -> (f32, f32)

where G: Module, D: Module, F: Fn(&Variable, &Variable) -> Variable,

Performs one adversarial training step.

Arguments

• generator - Generator model
• discriminator - Discriminator model
• real_data - Real data batch
• noise - Random noise input for generator
• loss_fn - Loss function taking (predictions, targets) → scalar loss

Returns

Tuple of (generator_loss, discriminator_loss)

pub fn optimizers(&self) -> (&dyn Optimizer, &dyn Optimizer)

Returns references to the optimizers.