def trainGA(self,
train_data,
target_data,
cf,
loss_fn=None,
verbose=False):
""" Train web with Genetic Algorithm """
train_data, target_data = self.check_cuda(train_data, target_data)
self.check_graph()
# prepare config object with information of web, # of genes, partitions, genomes, etc
self.prepare_config_obj(cf, loss_fn)
# initialize genepool
genepool = Evolution.GenePool(cf)
# stores which indices of self.parameters to change during training
self.set_dict_indices_from_pool(genepool.pool[0])
# np arrays to save genePools, outputs and fitness
geneArray = np.zeros((cf.generations, cf.genomes, cf.genes))
outputArray = np.zeros((cf.generations, cf.genomes,
train_data.shape[0], self.nr_output_vertices))
fitnessArray = np.zeros((cf.generations, cf.genomes))
# Temporary arrays, overwritten each generation
fitnessTemp = np.zeros((cf.genomes, cf.fitnessavg))
outputAvg = torch.zeros(cf.fitnessavg,
train_data.shape[0],
self.nr_output_vertices,
device=self.cuda)
outputTemp = torch.zeros(cf.genomes,
train_data.shape[0],
self.nr_output_vertices,
device=self.cuda)
for i in range(cf.generations):
for j in range(cf.genomes):
for avgIndex in range(cf.fitnessavg):
# update parameters of each network
self.set_parameters_from_pool(genepool.pool[j])
self.forward(train_data)
outputAvg[avgIndex] = self.get_output()
# use negative loss as fitness for genepool.NextGen()
fitnessTemp[j, avgIndex] = -cf.Fitness(
outputAvg[avgIndex], target_data).item()
outputTemp[j] = outputAvg[np.argmin(fitnessTemp[j])]
genepool.fitness = fitnessTemp.min(1) # Save best fitness
# Save generation data
geneArray[i, :, :] = genepool.pool
outputArray[i, :, :] = outputTemp
fitnessArray[i, :] = genepool.fitness
if verbose:
print("Generation nr. " + str(i + 1) + " completed")
print("Best fitness: " + str(-max(genepool.fitness)))
genepool.NextGen()
return geneArray, outputArray, fitnessArray
fitnessTemp = np.zeros((cf.genomes, cf.fitnessavg))
outputAvg = np.zeros((cf.fitnessavg, len(x[0])))
outputTemp = np.zeros((cf.genomes, len(x[0])))
controlVoltages = np.zeros(cf.genes)
# Initialize save directory
saveDirectory = SaveLib.createSaveDirectory(cf.filepath, cf.name)
# Initialize main figure
mainFig = PlotBuilder.initMainFigEvolution(cf.genes, cf.generations, cf.genelabels, cf.generange)
# Initialize instruments
ivvi = InstrumentImporter.IVVIrack.initInstrument()
# Initialize genepool
genePool = Evolution.GenePool(cf)
#%% Measurement loop
for i in range(cf.generations):
for j in range(cf.genomes):
# Set the DAC voltages
for k in range(cf.genes-1):
controlVoltages[k] = genePool.MapGenes(
cf.generange[k], genePool.pool[j, k])
InstrumentImporter.IVVIrack.setControlVoltages(ivvi, controlVoltages)
time.sleep(1) # Wait after setting DACs
# Set the input scaling
x_scaled = x * genePool.MapGenes(cf.generange[-1], genePool.pool[j, -1])
'''This test is for testing the Evolution module. The fitness function is
1/(1 + y.T*y), where y is a genome. In other words, the GA should arrive at
(unique) optimal solution y=0 with fitness = 1'''
import numpy as np
import matplotlib.pyplot as plt
import config_evolution_test as cf
import SkyNEt.modules.Evolution as evo
# Initialize config object
conf_obj = cf.experiment_config()
# Initialize genepool
pool = evo.GenePool(conf_obj)
# Initialize plotting variables
fitness = np.zeros((conf_obj.generations))
# Evolution loop
for i in range(conf_obj.generations):
for j in range(pool.genomes):
pool.fitness[j] = conf_obj.Fitness(pool.pool[j].T)
fitness[i] = max(pool.fitness) # Save fitness for plotting
print('Highest fitness of generation ' + str(i) + ': ' +
str(max(pool.fitness)))
print('Best genome of generation ' + str(i) + ': ' +
str(pool.pool[np.argmax(pool.fitness)]))
pool.NextGen()
passed_test = np.linalg.norm(pool.pool[0] - 0) < 1E-12