Generative Models

Entrenar provides Generative Adversarial Networks (GANs) for code synthesis.

Architecture

Latent Vector z ─┬─► Generator ─► AST Tokens ─┬─► Discriminator ─► Valid/Invalid
                 │                            │
                 │   Real AST Samples ────────┘
                 │
                 └── (sampled from N(0, I))

Key Components

Generator

• Maps latent vectors to Rust AST token sequences
• MLP architecture with configurable hidden layers
• Xavier/He initialization for stable training

Discriminator

• Classifies code as real (valid) or fake (invalid)
• Token embedding + MLP architecture
• Binary cross-entropy loss

Latent Space

• Standard normal distribution sampling
• SLERP interpolation for smooth transitions
• Supports latent space exploration

Quick Start

#![allow(unused)]
fn main() {
use entrenar::generative::{CodeGan, CodeGanConfig};

let config = CodeGanConfig::default();
let mut gan = CodeGan::new(config);

// Generate code samples
let latent = gan.sample_latent(10);
for z in &latent {
    let code = gan.generator.generate(z);
    println!("{:?}", code);
}
}

Use Cases

1. Code Completion: Generate likely next tokens
2. Data Augmentation: Create synthetic training data
3. Novelty Search: Explore code generation space
4. Transfer Learning: Pre-train on code distributions