def mutate(self, chromosome):
child_genes = np.copy(chromosome.genes)
change = np.random.rand(
len(child_genes)) < self.config.cfg["mutation_rate"]
child_genes[change] = np.random.rand(np.sum(change))
child = Chromosome(self.problem.nin, self.problem.nout)
child.from_genes(child_genes, self.config)
return child
def test_lamarckian():
clf.lamarckian = True
params = clf.get_params()
print("7 Params: ", params)
print("Node pids", ch.param_id)
print(len(ch.genes))
print("Before: ", ch.genes)
history = clf.train()
print("After: ", ch.genes)
params = clf.get_params()
acc = clf.evaluate()
print("8 Params: ", params)
print("Changed genes: ", ch.genes[orig_genes != ch.genes])
assert np.any(orig_genes != ch.genes)
ch2 = Chromosome(p.nin, p.nout)
ch2.from_genes(ch.genes, c)
clf2 = Classifier(ch2,
p.x_train,
p.x_test,
p.y_train,
p.y_test,
batch_size=p.batch_size,
epochs=p.epochs,
seed=p.seed,
lamarckian=p.lamarckian)
acc2 = clf.evaluate()
params2 = clf.get_params()
print("Accuracies: ", acc, acc2)
print("Params: ", params, params2)
print("Node pids: ", ch.param_id, ch2.param_id)
print("Genes: ", ch.genes, ch2.genes)
# assert acc == acc2
assert np.all(params == params2)
assert ch.param_id == ch2.param_id
for nid in range(len(ch.nodes)):
assert ch.nodes[nid].param == ch2.nodes[nid].param
class LearnEvolver:
def __init__(self, problem, config, logname='test', root_dir='.'):
self.config = config
self.problem = problem
self.epochs = 1 * self.problem.epochs
self.max_learn_fit = 0.0
self.max_evo_fit = 0.0
self.max_fit = 0.0
self.evo_best = Chromosome(self.problem.nin, self.problem.nout)
self.evo_best.random(config)
self.learn_best = Chromosome(self.problem.nin, self.problem.nout)
self.learn_best.from_genes(self.evo_best.genes, config)
self.logfile = os.path.join(root_dir, 'logs', logname + '.log')
self.logname = logname
self.generation = 0
def run(self, n_steps):
for i in range(n_steps):
self.step()
def mutate(self, chromosome):
child_genes = np.copy(chromosome.genes)
change = np.random.rand(
len(child_genes)) < self.config.cfg["mutation_rate"]
child_genes[change] = np.random.rand(np.sum(change))
child = Chromosome(self.problem.nin, self.problem.nout)
child.from_genes(child_genes, self.config)
return child
def step(self):
evo_delete = True
learn_delete = True
evo_child = self.mutate(self.evo_best)
self.problem.epochs = 0
fitness, history = self.problem.get_fitness(evo_child)
if fitness >= self.max_evo_fit:
self.evo_best = evo_child
self.max_evo_fit = fitness
evo_delete = False
if fitness >= self.max_learn_fit:
self.learn_best = evo_child
self.max_learn_fit = fitness
evo_delete = False
learn_child = self.mutate(self.learn_best)
self.problem.epochs = self.epochs
fitness, history = self.problem.get_fitness(learn_child)
if fitness >= self.max_learn_fit:
self.learn_best = learn_child
self.max_learn_fit = fitness
learn_delete = False
new_max = max(self.max_evo_fit, self.max_learn_fit)
if new_max > self.max_fit:
self.max_fit = new_max
with open(self.logfile, 'a') as f:
for i in range(len(history)):
# type,logname,gen,eval,epoch,total_epochs,loss,acc,best
f.write('L,%s,%d,%d,%d,%d,%0.10f,%0.10f,%0.10f\n' %
(self.logname, self.generation,
self.problem.eval_count, i, self.problem.epochs,
history[i][0], history[i][1], new_max))
f.write(
'E,%s,%d,%d,%d,%d,%0.10f,%0.10f,%0.10f\n' %
(self.logname, self.generation, self.problem.eval_count, 0,
self.problem.epochs, 0.0, 0.0, new_max))
if evo_delete:
del evo_child
if learn_delete:
del learn_child
self.generation += 1
