rustytorch/crates/training/rtx-transformers/examples/fomaml_demo.rs

//! FOMAML (First-Order Model-Agnostic Meta-Learning) Demonstration
//!
//! This example demonstrates the complete FOMAML pipeline for few-shot learning,
//! including task sampling, inner loop adaptation, and meta-parameter updates.

use rtx_tensor::Device;
use rtx_transformers::meta::*;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    println!("🚀 FOMAML Demo - First-Order Model-Agnostic Meta-Learning");
    println!("=========================================================");

    // Setup device
    let device = Device::cuda(0).unwrap_or(Device::default());
    println!("📱 Using device: {:?}", device);

    // FOMAML Configuration
    let fomaml_config = FOMAMLConfig {
        inner_lr: 0.01,         // Inner loop learning rate
        outer_lr: 0.001,        // Meta learning rate
        num_inner_steps: 5,     // Adaptation steps
        num_tasks_per_batch: 4, // Tasks per meta-update
        first_order: true,      // First-order approximation (key feature of FOMAML)
        num_ways: 5,            // 5-way classification
        num_shots: 1,           // 1-shot learning
        num_queries: 15,        // Query set size
    };

    println!("⚙️ FOMAML Configuration:");
    println!("   - Inner LR: {}", fomaml_config.inner_lr);
    println!("   - Outer LR: {}", fomaml_config.outer_lr);
    println!("   - Inner steps: {}", fomaml_config.num_inner_steps);
    println!(
        "   - Tasks per batch: {}",
        fomaml_config.num_tasks_per_batch
    );
    println!("   - First-order: {}", fomaml_config.first_order);
    println!(
        "   - N-way: {}, K-shot: {}",
        fomaml_config.num_ways, fomaml_config.num_shots
    );

    // Validate configuration
    fomaml_config.validate()?;
    println!("✅ Configuration validated successfully");

    // Create synthetic few-shot dataset
    let num_classes = 10;
    let samples_per_class = 20;
    let feature_dim = 64;

    let dataset = FewShotDataset::synthetic(num_classes, samples_per_class, feature_dim, &device)?;
    println!("📊 Created synthetic dataset:");
    println!("   - Classes: {}", num_classes);
    println!("   - Samples per class: {}", samples_per_class);
    println!("   - Feature dimension: {}", feature_dim);

    // Initialize FOMAML model
    let hidden_dim = 32;
    let output_dim = fomaml_config.num_ways;

    let mut fomaml = FOMAML::new(
        feature_dim,
        hidden_dim,
        output_dim,
        fomaml_config.clone(),
        &device,
    )?;
    println!("🧠 Initialized FOMAML model:");
    println!("   - Input dim: {}", feature_dim);
    println!("   - Hidden dim: {}", hidden_dim);
    println!("   - Output dim: {}", output_dim);

    // Create task sampler
    let sampler_config = TaskSamplerConfig {
        n_way: fomaml_config.num_ways,
        k_shot: fomaml_config.num_shots,
        query_per_class: fomaml_config.num_queries / fomaml_config.num_ways,
        seed: Some(42),
    };

    let sampler = TaskSampler::new(dataset, sampler_config)?;
    println!(
        "🎯 Created task sampler for {}-way {}-shot learning",
        fomaml_config.num_ways, fomaml_config.num_shots
    );

    // Training phase
    println!("\n🏋️ Training Phase");
    println!("================");

    let num_meta_updates = 10;
    for meta_update in 1..=num_meta_updates {
        // Sample tasks for this meta-batch
        let tasks = sampler.sample_fomaml_tasks(&fomaml_config, &device)?;

        // Perform FOMAML meta-update
        let stats = fomaml.meta_update(tasks, &device)?;

        if meta_update % 2 == 0 {
            println!(
                "Meta-update {}/{}: Loss={:.4}, Accuracy={:.2}%",
                meta_update,
                num_meta_updates,
                stats.avg_outer_loss,
                stats.avg_query_accuracy * 100.0
            );
        }
    }

    let final_stats = fomaml.get_stats();
    println!("📈 Final Training Statistics:");
    println!("   - Meta-updates: {}", final_stats.meta_updates);
    println!("   - Tasks processed: {}", final_stats.tasks_processed);
    println!("   - Avg inner loss: {:.4}", final_stats.avg_inner_loss);
    println!("   - Avg outer loss: {:.4}", final_stats.avg_outer_loss);
    println!(
        "   - Avg accuracy: {:.2}%",
        final_stats.avg_query_accuracy * 100.0
    );

    // Evaluation phase
    println!("\n🎯 Evaluation Phase");
    println!("==================");

    let num_test_episodes = 5;
    let mut test_accuracies = Vec::new();
    let mut test_losses = Vec::new();

    for episode in 1..=num_test_episodes {
        // Sample a test task
        let test_task = sampler.sample_task(&device)?;

        // Fast adaptation and evaluation
        let (accuracy, loss) = fomaml.evaluate_few_shot(&test_task, &device)?;

        test_accuracies.push(accuracy);
        test_losses.push(loss);

        println!(
            "Test episode {}: Accuracy={:.2}%, Loss={:.4}",
            episode,
            accuracy * 100.0,
            loss
        );
    }

    let mean_accuracy = test_accuracies.iter().sum::<f32>() / test_accuracies.len() as f32;
    let mean_loss = test_losses.iter().sum::<f32>() / test_losses.len() as f32;

    println!("\n📊 Test Results Summary:");
    println!("   - Mean accuracy: {:.2}%", mean_accuracy * 100.0);
    println!("   - Mean loss: {:.4}", mean_loss);

    // Demonstrate key FOMAML features
    println!("\n🔍 FOMAML Key Features Demonstrated:");
    println!("====================================");
    println!("✅ First-order approximation (computational efficiency)");
    println!("✅ Task sampling for N-way K-shot learning");
    println!("✅ Inner loop adaptation with gradient detachment");
    println!("✅ Outer loop meta-parameter updates");
    println!("✅ Few-shot learning evaluation");
    println!("✅ Batch processing of multiple tasks");

    // Performance comparison note
    println!("\n⚡ Performance Benefits of FOMAML:");
    println!("- Faster than full MAML (no second derivatives)");
    println!("- More memory efficient");
    println!("- Suitable for larger models and more adaptation steps");
    println!("- Maintains competitive few-shot learning performance");

    println!("\n🎉 FOMAML demo completed successfully!");

    Ok(())
}