Struct MemoryEfficientLayer

pub struct MemoryEfficientLayer { /* private fields */ }

Memory-efficient neural network layer that processes data in chunks

Implementations

impl MemoryEfficientLayer

pub fn new( input_size: usize, output_size: usize, chunk_size: Option<usize>, ) -> Result<Self>

Create a new memory-efficient layer

Examples found in repository

examples/memory_efficient_example.rs (lines 228-232)

fn demo_memory_efficient_layer() -> Result<()> {
    println!("\n🧠 Memory-Efficient Layer Demo");
    println!("------------------------------");

    // Create a memory-efficient layer
    let layer = MemoryEfficientLayer::new(
        784,      // Input size (e.g., 28x28 MNIST)
        128,      // Output size
        Some(64), // Chunk size
    )?;

    println!("Created memory-efficient layer: 784 -> 128");

    // Create input data
    let input =
        Array2::from_shape_fn((256, 784), |(i, j)| ((i + j) as f32 * 0.01).sin()).into_dyn();

    println!("Input shape: {:?}", input.shape());

    // Forward pass
    println!("Performing forward pass...");
    let start = Instant::now();
    let output = layer.forward(&input)?;
    let forward_time = start.elapsed();

    println!("Forward pass completed in {:?}", forward_time);
    println!("Output shape: {:?}", output.shape());

    // Verify output statistics
    let mean = output.mean().unwrap_or(0.0);
    let std = {
        let variance = output.mapv(|x| (x - mean).powi(2)).mean().unwrap_or(0.0);
        variance.sqrt()
    };

    println!("Output statistics:");
    println!("  Mean: {:.6}", mean);
    println!("  Std: {:.6}", std);
    println!(
        "  Min: {:.6}",
        output.iter().cloned().fold(f32::INFINITY, f32::min)
    );
    println!(
        "  Max: {:.6}",
        output.iter().cloned().fold(f32::NEG_INFINITY, f32::max)
    );

    Ok(())
}

pub fn forward(&self, input: &ArrayD<f32>) -> Result<ArrayD<f32>>

Forward pass with memory-efficient chunk processing

Examples found in repository

examples/memory_efficient_example.rs (line 245)

fn demo_memory_efficient_layer() -> Result<()> {
    println!("\n🧠 Memory-Efficient Layer Demo");
    println!("------------------------------");

    // Create a memory-efficient layer
    let layer = MemoryEfficientLayer::new(
        784,      // Input size (e.g., 28x28 MNIST)
        128,      // Output size
        Some(64), // Chunk size
    )?;

    println!("Created memory-efficient layer: 784 -> 128");

    // Create input data
    let input =
        Array2::from_shape_fn((256, 784), |(i, j)| ((i + j) as f32 * 0.01).sin()).into_dyn();

    println!("Input shape: {:?}", input.shape());

    // Forward pass
    println!("Performing forward pass...");
    let start = Instant::now();
    let output = layer.forward(&input)?;
    let forward_time = start.elapsed();

    println!("Forward pass completed in {:?}", forward_time);
    println!("Output shape: {:?}", output.shape());

    // Verify output statistics
    let mean = output.mean().unwrap_or(0.0);
    let std = {
        let variance = output.mapv(|x| (x - mean).powi(2)).mean().unwrap_or(0.0);
        variance.sqrt()
    };

    println!("Output statistics:");
    println!("  Mean: {:.6}", mean);
    println!("  Std: {:.6}", std);
    println!(
        "  Min: {:.6}",
        output.iter().cloned().fold(f32::INFINITY, f32::min)
    );
    println!(
        "  Max: {:.6}",
        output.iter().cloned().fold(f32::NEG_INFINITY, f32::max)
    );

    Ok(())
}