User Guide

Learn how to use Eggroll Trainer effectively.

Sections

• Core Concepts - Understanding Evolution Strategies and EGGROLL
• Trainers - Using ESTrainer and EGGROLLTrainer
• Fitness Functions - Writing effective fitness functions
• Advanced Usage - Custom algorithms, device management, tuning

Overview

Eggroll Trainer provides three main classes:

1. ESTrainer - Base class for custom ES algorithms
2. VanillaESTrainer - Vanilla ES with full-rank perturbations
3. EGGROLLTrainer - Advanced ES with low-rank perturbations (recommended)

Most users should start with EGGROLLTrainer for best performance.

Quick Reference

Basic Training Loop

from eggroll_trainer import EGGROLLTrainer

trainer = EGGROLLTrainer(
    model.parameters(),
    model=model,
    fitness_fn=fitness_fn,
    population_size=256,
    learning_rate=0.01,
    sigma=0.1,
)

trainer.train(num_generations=100)
best_model = trainer.get_best_model()

Accessing Training History

# After training
history = trainer.history
print(f"Best fitness: {max(history['fitness'])}")
print(f"Mean fitness over time: {history['mean_fitness']}")

Custom ES Algorithm

from eggroll_trainer import ESTrainer

class MyESTrainer(ESTrainer):
    def sample_perturbations(self, population_size):
        # Your perturbation sampling logic
        pass

    def compute_update(self, perturbations, fitnesses):
        # Your update computation logic
        pass

Next Steps

Start with Core Concepts to understand the fundamentals.