Examples

After installing the required dependencies as outlined in the requirements file, yyou can utilize genetic algorithms for hyperparameter tuning in regression tasks using both tensorflow.keras and torch models. For example, the GitHub repository contains examples that demonstrate neural network reconstruction of the distance modulus using the Joint Light Analysis (JLA) compilation, similar to the IV.A section of the reference Gómez-Vargas, I., Andrade, J. B., & Vázquez, J. A. (2023). Neural networks optimized by genetic algorithms in cosmology. Physical Review D, 107(4), 043509..

Using tensorflow.keras:

import time
from nnogada.Nnogada import Nnogada
import pandas as pd
from sklearn.preprocessing import StandardScaler
import numpy as np


df = pd.read_csv('data/jla.csv')
N = len(df.values)
randomize = np.random.permutation(N)
data = df.values[randomize]
N = len(df.values)
z = data[:, 0]
y = data[:, 1:3] ### toma el resto de variables a predecir
y[:,1] = y[:, 1]**2+data[:,2]
np.shape(y)


dmag = df["dmb"]
df2 = df['errors']+df['dmb']**2

scalerz = StandardScaler()
scalerz.fit(z.reshape(-1, 1))
z = scalerz.transform(z.reshape(-1, 1))

split = 0.75
ntrain = int(split * len(z))
indx = [ntrain]
X_train, X_val = np.split(z, indx)
Y_train, Y_val = np.split(y, indx)

population_size = 11   # max of individuals per generation
max_generations = 5    # number of generations
gene_length = 4        # lenght of the gene, depends on how many hiperparameters are tested
k = 1                  # num. of finalist individuals

t = time.time()
datos = []

# Define the hyperparameters for the search
hyperparams = {'deep': [1, 2], 'num_units': [1,2], 'batch_size': [256,1048]}

# generate a Nnogada instance
net_fit = Nnogada(hyp_to_find=hyperparams, X_train=X_train, Y_train=Y_train, X_val=X_val, Y_val=Y_val)
# Set the possible values of hyperparameters and not use the default values from hyperparameters.py
net_fit.set_hyperparameters()

# best solution
best_population = net_fit.ga_with_elitism(population_size, max_generations, gene_length, k)
print(best_population)
print("Total elapsed time:", (time.time()-t)/60, "minutes")

Using torch:

import time
from nnogada.Nnogada import Nnogada
import pandas as pd
from sklearn.preprocessing import StandardScaler
import numpy as np


df = pd.read_csv('data/jla.csv')
N = len(df.values)
randomize = np.random.permutation(N)
data = df.values[randomize]
N = len(df.values)
z = data[:, 0]
y = data[:, 1:3] ### toma el resto de variables a predecir
y[:,1] = y[:, 1]**2+data[:,2]
np.shape(y)


dmag = df["dmb"]
df2 = df['errors']+df['dmb']**2

scalerz = StandardScaler()
scalerz.fit(z.reshape(-1, 1))
z = scalerz.transform(z.reshape(-1, 1))

split = 0.75
ntrain = int(split * len(z))
indx = [ntrain]
X_train, X_val = np.split(z, indx)
Y_train, Y_val = np.split(y, indx)

population_size = 11   # max of individuals per generation
max_generations = 5    # number of generations
gene_length = 4        # lenght of the gene, depends on how many hiperparameters are tested
k = 1                  # num. of finalist individuals

t = time.time()
# Define the hyperparameters for the search
hyperparams = {'deep': [1, 2], 'num_units': [1,2], 'batch_size': [2048,1048]}

# generate a Nnogada instance
net_fit = Nnogada(hyp_to_find=hyperparams, X_train=X_train, Y_train=Y_train, X_val=X_val, Y_val=Y_val,
                  neural_library='torch')
# Set the possible values of hyperparameters and not use the default values from hyperparameters.py
net_fit.set_hyperparameters()

# Find best solutions
net_fit.ga_with_elitism(population_size, max_generations, gene_length, k)
# best solution
print("best individual", net_fit.best)
print("Best number of nodes:", net_fit.best['num_units'])
print("Best number of layers:", net_fit.best['deep'])
print("Best number of batch_size:", net_fit.best['batch_size'])
print("Total elapsed time:", (time.time()-t)/60, "minutes")