Skill 050

Requirements for Outputs

General Guidelines

Optimization

• All optimization routines MUST support JAX-compatible parameters and gradients.
• Ensure numerical stability and convergence for all optimization methods.
• Provide informative error messages for invalid inputs or configurations.

JAX Optimization Techniques

1. Gradient Descent

gradient_descent(fn, init_params, learning_rate, num_steps)

Description: Perform gradient descent optimization on a loss function. Parameters:

• fn (callable): The loss function to minimize.
• init_params (array): Initial parameters for optimization.
• learning_rate (float): Step size for each iteration.
• num_steps (int): Number of optimization steps.

Returns: Optimized parameters after the specified number of steps.

import jax
import jax.numpy as jnp
from jax import grad
import jax_optimization_techniques as jx

# Define a simple quadratic loss function
def loss_fn(params):
    return jnp.sum((params - 3) ** 2)

# Perform optimization
optimized_params = jx.gradient_descent(loss_fn, jnp.array([0.0]), learning_rate=0.1, num_steps=100)

2. Adam Optimizer

adam_optimizer(fn, init_params, learning_rate, num_steps)

Description: Optimize a loss function using the Adam optimization algorithm. Parameters:

• fn (callable): The loss function to minimize.
• init_params (array): Initial parameters for optimization.
• learning_rate (float): Step size for each iteration.
• num_steps (int): Number of optimization steps.

Returns: Optimized parameters after the specified number of steps.

# Initialize Adam optimizer
def adam_optimizer(fn, init_params, learning_rate=0.001, num_steps=100):
    params = init_params
    m = jax.numpy.zeros_like(params)
    v = jax.numpy.zeros_like(params)
    beta1 = 0.9
    beta2 = 0.999
    epsilon = 1e-8

    for t in range(1, num_steps + 1):
        g = grad(fn)(params)
        m = beta1 * m + (1 - beta1) * g
        v = beta2 * v + (1 - beta2) * (g ** 2)
        m_hat = m / (1 - beta1 ** t)
        v_hat = v / (1 - beta2 ** t)
        params -= learning_rate * m_hat / (jnp.sqrt(v_hat) + epsilon)
    return params

optimized_params = adam_optimizer(loss_fn, jnp.array([0.0]), learning_rate=0.01, num_steps=100)

3. Learning Rate Schedulers

exponential_decay(initial_lr, global_step, decay_steps, decay_rate)

Description: Calculate the learning rate at a given step using exponential decay. Parameters:

• initial_lr (float): Initial learning rate.
• global_step (int): Current training step.
• decay_steps (int): Step interval for decay.
• decay_rate (float): Rate of decay.

Returns: Adjusted learning rate.

def exponential_decay(initial_lr, global_step, decay_steps, decay_rate):
    return initial_lr * (decay_rate ** (global_step // decay_steps))

current_lr = exponential_decay(0.1, 50, 10, 0.96)

4. Hyperparameter Tuning

hyperparameter_tuning(fn, param_grid)

Description: Perform hyperparameter tuning to find optimal parameters for a model. Parameters:

• fn (callable): The model training function that takes hyperparameters as input.
• param_grid (dict): Dictionary of hyperparameters and their corresponding list of values to test.

Returns: Best parameters based on validation performance.

from sklearn.model_selection import ParameterGrid

def tune_hyperparameters(fn, param_grid):
    best_score = float('inf')
    best_params = None
    for params in ParameterGrid(param_grid):
        score = fn(**params)
        if score < best_score:
            best_score = score
            best_params = params
    return best_params

best_params = tune_hyperparameters(train_model, {'learning_rate': [0.01, 0.1], 'batch_size': [32, 64]})

Notes:

• Ensure that the loss function and tuning parameters are well-defined.
• Use cross-validation for better evaluation of model performance.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: legendtkl
• Source: legendtkl/agentic-skill-router
• License: MIT
• Homepage: https://legendtkl.github.io/agentic-skill-router/

Install and usage instructions live in the source repository linked above.