def run():
# load the config file
local_dir = os.path.dirname(__file__)
config_path = os.path.join(local_dir, my_config)
config = neat.Config(AgentGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
config_path)
# uses the extended NEAT population PopulationSyn that synchronizes with singularity
pop = neat.Population(config)
# add reporters
stats = neat.StatisticsReporter()
pop.add_reporter(stats)
pop.add_reporter(neat.StdOutReporter(True))
# save a checkpoint every 100 generations or 900 seconds.
rep = neat.Checkpointer(100, 900)
pop.add_reporter(rep)
# class for evaluating the population
ec = GenomeEvaluator(ts_f, vs_f)
# initializes genomes fitness and gen_best just for the first time
for g in itervalues(pop.population):
g.fitness = -10000000.0
ec.genomes_h.append(g)
gen_best = g
# initializations
avg_score_v = -10000000.0
avg_score_v_ant = avg_score_v
avg_score = avg_score_v
iteration_counter = 0
best_fitness = -2000.0
pop_size = len(pop.population)
# sets the nuber of continuous iterations
num_iterations = round(200 / len(pop.population)) + 1
# repeat NEAT iterations until solved or keyboard interrupt
while 1:
try:
# if it is not the first iteration calculate training and validation scores
if iteration_counter > 0:
avg_score = ec.training_validation_score(gen_best, config)
# if it is not the first iteration
if iteration_counter >= 0:
# synchronizes with singularity migrating maximum 3 specimens
# pop.syn_singularity(4, my_url, stats,avg_score,rep.current_generation, config, ec.genomes_h)
pop.species.speciate(config, pop.population, pop.generation)
print("\nSpeciation after migration done")
# perform pending evaluations on the singularity network, max 2
#pop.evaluate_pending(2)
#increment iteration counter
iteration_counter = iteration_counter + 1
# execute num_iterations consecutive iterations of the NEAT algorithm
gen_best = pop.run(ec.evaluate_genomes, 2)
# verify the training score is enough to stop the NEAT algorithm: TODO change to validation score when generalization is ok
if avg_score < 2000000000:
solved = False
if solved:
print("Solved.")
# save the winners.
for n, g in enumerate(best_genomes):
name = 'winner-{0}'.format(n)
with open(name + '.pickle', 'wb') as f:
pickle.dump(g, f)
break
except KeyboardInterrupt:
print("User break.")
break
env.close()
def run(config_file):
# Load configuration.
config = neat.Config(
neat.DefaultGenome,
neat.DefaultReproduction,
neat.DefaultSpeciesSet,
neat.DefaultStagnation,
config_file,
)
# Create the population, which is the top-level object for a NEAT run.
p = neat.Population(config)
# Add a stdout reporter to show progress in the terminal.
p.add_reporter(neat.StdOutReporter(True))
stats = neat.StatisticsReporter()
p.add_reporter(stats)
# Run for up to 300 generations.
pe = neat.ParallelEvaluator(multiprocessing.cpu_count(), eval_genome)
winner = p.run(pe.evaluate, runs)
winner_net = neat.nn.FeedForwardNetwork.create(winner, config)
outputs = []
for xi, xo in zip(x_inputs, x_outputs):
output = winner_net.activate(xi)
outputs.append(output)
print("input {!r}, expected output {!r}, got {!r}".format(
xi, xo, output))
xi = tuple([0, 0])
xo = [tuple([0])]
output = winner_net.activate(xi)
print("input {!r}, expected output {!r}, got {!r}".format(xi, xo, output))
visualize.draw_net(config, winner, True)
visualize.plot_stats(stats, ylog=False, view=True)
visualize.plot_species(stats, view=True)
fig = plt.figure(figsize=(16, 9))
ax = fig.gca(projection='3d')
# Make data.
X = np.arange(-32.768, 32.768, 0.25)
Y = np.arange(-32.768, 32.768, 0.25)
X, Y = np.meshgrid(X, Y)
OZ = []
for i in range(X.shape[0]):
for j in range(X.shape[1]):
a = np.array([X[i][j] / 32.768, Y[i][j] / 32.768])
OZ.append(winner_net.activate(tuple(a)))
OZ = np.array(OZ).reshape(X.shape)
Z = OZ * 24
# Plot the surface.
surf = ax.plot_surface(X,
Y,
Z,
cmap=cm.coolwarm,
linewidth=0,
antialiased=False,
alpha=0.4)
# Customize the z axis.
ax.set_zlim(-1.01, 25)
ax.zaxis.set_major_locator(LinearLocator(10))
ax.zaxis.set_major_formatter(FormatStrFormatter('%.02f'))
outputs = []
for xi in x_inputs:
output = winner_net.activate(xi)
outputs.append(output)
x = []
y = []
z = []
for i in range(1000):
x.append(x_inputs[i][0] * 32.768)
y.append(x_inputs[i][1] * 32.768)
z.append(outputs[i][0] * 24)
ax.scatter(x, y, z, c='k', marker='o')
# Add a color bar which maps values to colors.
fig.colorbar(surf, shrink=0.5, aspect=5)
plt.savefig("fitting_ackley2")
return winner
def prescribe(start_date_str: str, end_date_str: str,
path_to_prior_ips_file: str, path_to_cost_file: str,
output_file_path) -> None:
start_date = pd.to_datetime(start_date_str, format='%Y-%m-%d')
end_date = pd.to_datetime(end_date_str, format='%Y-%m-%d')
# Load historical data with basic preprocessing
print("Loading historical data...")
df = prepare_historical_df()
# Restrict it to dates before the start_date
df = df[df['Date'] <= start_date]
# Fill in any missing case data using predictor given ips_df.
# todo: ignore ips_df for now, and instead assume we have case
# data for all days and geos up until the start_date.
# Create historical data arrays for all geos
past_cases = {}
past_ips = {}
for geo in df['GeoID'].unique():
geo_df = df[df['GeoID'] == geo]
past_cases[geo] = np.maximum(0, np.array(geo_df[CASES_COL]))
past_ips[geo] = np.array(geo_df[IP_COLS])
# Gather values for scaling network output
ip_max_values_arr = np.array([IP_MAX_VALUES[ip] for ip in IP_COLS])
# Load prescriptors
checkpoint = neat.Checkpointer.restore_checkpoint(PRESCRIPTORS_FILE)
prescriptors = checkpoint.population.values()
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
'config-prescriptor')
# Load IP costs to condition prescriptions
cost_df = pd.read_csv(path_to_cost_file)
cost_df['RegionName'] = cost_df['RegionName'].fillna("")
cost_df = add_geo_id(cost_df)
geo_costs = {}
for geo in cost_df['GeoID'].unique():
costs = cost_df[cost_df['GeoID'] == geo]
cost_arr = np.array(costs[IP_COLS])[0]
geo_costs[geo] = cost_arr
# Generate prescriptions
prescription_dfs = []
for prescription_idx, prescriptor in enumerate(prescriptors):
print("Generating prescription", prescription_idx, "...")
# Create net from genome
net = neat.nn.FeedForwardNetwork.create(prescriptor, config)
# Set up dictionary for keeping track of prescription
df_dict = {'CountryName': [], 'RegionName': [], 'Date': []}
for ip_col in sorted(IP_MAX_VALUES.keys()):
df_dict[ip_col] = []
# Set initial data
eval_past_cases = deepcopy(past_cases)
eval_past_ips = deepcopy(past_ips)
# Generate prescriptions one day at a time, feeding resulting
# predictions from the predictor back into the prescriptor.
for date in pd.date_range(start_date, end_date):
date_str = date.strftime("%Y-%m-%d")
# Get prescription for all regions
for geo in df['GeoID'].unique():
# Prepare input data. Here we use log to place cases
# on a reasonable scale; many other approaches are possible.
X_cases = np.log(eval_past_cases[geo][-NB_LOOKBACK_DAYS:] + 1)
X_ips = eval_past_ips[geo][-NB_LOOKBACK_DAYS:]
X_costs = geo_costs[geo]
X = np.concatenate(
[X_cases.flatten(),
X_ips.flatten(), X_costs])
# Get prescription
prescribed_ips = net.activate(X)
# Map prescription to integer outputs
prescribed_ips = (prescribed_ips * ip_max_values_arr).round()
# Add it to prescription dictionary
country_name, region_name = geo.split('__')
if region_name == 'nan':
region_name = np.nan
df_dict['CountryName'].append(country_name)
df_dict['RegionName'].append(region_name)
df_dict['Date'].append(date_str)
for ip_col, prescribed_ip in zip(IP_COLS, prescribed_ips):
df_dict[ip_col].append(prescribed_ip)
# Create dataframe from prescriptions
pres_df = pd.DataFrame(df_dict)
# Make prediction given prescription for all countries
pred_df = get_predictions(start_date_str, date_str, pres_df)
# Update past data with new day of prescriptions and predictions
pres_df['GeoID'] = pres_df['CountryName'] + '__' + pres_df[
'RegionName'].astype(str)
pred_df['RegionName'] = pred_df['RegionName'].fillna("")
pred_df['GeoID'] = pred_df['CountryName'] + '__' + pred_df[
'RegionName'].astype(str)
new_pres_df = pres_df[pres_df['Date'] == date_str]
new_pred_df = pred_df[pred_df['Date'] == date_str]
for geo in df['GeoID'].unique():
geo_pres = new_pres_df[new_pres_df['GeoID'] == geo]
geo_pred = new_pred_df[new_pred_df['GeoID'] == geo]
# Append array of prescriptions
pres_arr = np.array([
geo_pres[ip_col].values[0] for ip_col in IP_COLS
]).reshape(1, -1)
eval_past_ips[geo] = np.concatenate(
[eval_past_ips[geo], pres_arr])
# It is possible that the predictor does not return values for some regions.
# To make sure we generate full prescriptions, this script continues anyway.
# Geos that are ignored in this way by the predictor, will not be used in
# quantitative evaluation. A list of such geos can be found in unused_geos.txt.
if len(geo_pred) != 0:
eval_past_cases[geo] = np.append(
eval_past_cases[geo],
geo_pred[PRED_CASES_COL].values[0])
# Add prescription df to list of all prescriptions for this submission
pres_df['PrescriptionIndex'] = prescription_idx
prescription_dfs.append(pres_df)
# Combine dfs for all prescriptions into a single df for the submission
prescription_df = pd.concat(prescription_dfs)
# Create the output path
os.makedirs(os.path.dirname(output_file_path), exist_ok=True)
# Save to a csv file
prescription_df.to_csv(output_file_path, index=False)
print('Prescriptions saved to', output_file_path)
return
#Loading classifiers
lexiwinner_vn = load_object('lexiwinner_vn')
lexiwinner_n = load_object('lexiwinner_n')
lexiwinner_neu = load_object('lexiwinner_neu')
lexiwinner_p = load_object('lexiwinner_p')
lexiwinner_vp = load_object('lexiwinner_vp')
#Loading test_inputs
with open('lexitest_inputs.txt', 'rb') as f:
test_inputs = pickle.load(f)
#Creating neural networks from winning genomes
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
'settings.ini')
net_vn = neat.nn.FeedForwardNetwork.create(lexiwinner_vn, config)
net_n = neat.nn.FeedForwardNetwork.create(lexiwinner_n, config)
net_neu = neat.nn.FeedForwardNetwork.create(lexiwinner_neu, config)
net_p = neat.nn.FeedForwardNetwork.create(lexiwinner_p, config)
net_vp = neat.nn.FeedForwardNetwork.create(lexiwinner_vp, config)
#Creating solution set
solution = pd.DataFrame({
'vn': [net_vn.activate(x)[0] for x in test_inputs],
'n': [net_n.activate(x)[0] for x in test_inputs],
'neu': [net_neu.activate(x)[0] for x in test_inputs],
'p': [net_p.activate(x)[0] for x in test_inputs],
max_episode_steps=1000,
tags={"pg_complexity": 8000000})
def int_a(a):
return np.array(a)
def fitness(rec):
f = rec['reward'].sum(axis=1).mean()
return f
# Load configuration
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
os.path.join(os.path.dirname(__file__), 'config-miko'))
# Construct experiment
exp = NEATGymExperiment(
'MiKo-v1',
config,
interpret_action=int_a,
runs_per_genome=2,
extract_fitness=fitness,
mode='parallel',
instances=7,
# render_all=True,
# network=neat.nn.GRUNetwork,
# starting_gen=0
)
def run(config_file):
# carrega o arquivo de configuracao realizando os "parsers" necessarios para coletar as configuracoes.
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
config_file)
# Cria uma populacao de genomas baseadas nas configuracoes setadas no arquivo
p = neat.Population(config)
# Adiciona algumas configuracoes de saidas no terminal para gerar pequenos relatorios durante a execucaoo.
p.add_reporter(neat.StdOutReporter(True))
stats = neat.StatisticsReporter()
p.add_reporter(stats)
# informa no console e salva um arquivo com o estado atual da populacao de genotipos atual para evitar
# perdas durante a execucao caso ocorra uma falha de energia por exemplo.
p.add_reporter(neat.Checkpointer(5))
# Executa N vezes o metodo de qualificacao de fitness para toda a populacao parametro 1 passa-se o metodo
# e o segundo a quantidade de geracoes em que sera executada. Caso atinja-se o fitness antes do valor de
# geracoes estimado a execucao e interrompida visto que o objetivo primario e sempre atingir o valor de fitness desejado
winner = p.run(eval_genomes, 100)
# Mostra qual o genoma obteve o maior fitness dentre as N execucoes de avaliacao e ou o melhor genoma que atingiu o valor
# igual ou superior ou desejado e determinado no arquivo de configuracao.
print('\nBest genome:\n{!s}'.format(winner))
# Transforma o genoma e fenotipo para uma analise visual do resultado dele
print('\nOutput:')
winner_net = neat.nn.RecurrentNetwork.create(winner, config)
# lista com os nomes das entradas da rede neural com os valores negativos
# lista com os nomes das saidas da rede neural com os valores positivos
node_names = {
-1: 'dintancia projetil 1',
-2: 'distancia projetil 2',
-3: 'distancia projetil 3',
-4: 'distancia projetil 4',
-5: 'distancia projetil 5',
-6: 'distancia projetil 6',
-7: 'distancia projetil 7',
-8: 'distancia projetil 8',
-9: 'distancia projetil 9',
-10: 'distancia projetil 10',
-11: 'distancia projetil 11',
-12: 'distancia projetil 12',
-13: 'distancia projetil 13',
-14: 'distancia projetil 14',
-15: 'distancia projetil 15',
-16: 'distancia projetil 16',
-17: 'distancia inimigo 1',
-18: 'distancia inimigo 2',
-19: 'direcao sprites 1',
-20: 'direcao sprites 2',
0: 'esquerda',
1: 'direita',
2: 'pular',
3: 'atirar',
4: 'release'
}
# desenha uma representacao visual do fenotipo "campeao"
visualize.draw_net(config, winner, True, node_names=node_names)
# desenha um grafico de duas dimensoes com y sendo o fitness e x sendo as geracoes,
# demonstrando assim a evolucao do nivel de fitness pelo passar das geracoes
visualize.plot_stats(stats, ylog=False, view=True)
# desenha visualmente a variacao e geracao de novas especies com o passar das geracoes
visualize.plot_species(stats, view=True)
counter += 1
# Si la partida acaba (mono colisiona con Charlie)
if done:
# Marca la partida como terminada
done = True
# Despliega el numero del genome con su puntaje obtenido
print(genome_id, current_score)
# Actualiza el puntaje del genoma por cada iteración de este
genome.fitness = current_score
if __name__ == "__main__":
# Inicializa la configuración de la red neuronal y las caracteristicas de los genomas
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
'config-feedforward')
p = neat.Population(config)
p.add_reporter(neat.StdOutReporter(True))
p.add_reporter(neat.StatisticsReporter())
p.add_reporter(neat.Checkpointer(10))
winner = p.run(eval_genomes)
with open('winner.pkl', 'wb') as output:
pickle.dump(winner, output, 1)
g.graph_attr.update(splines="false", nodesep='1', ranksep='2')
with g.subgraph(name="input") as c:
for n in layers['input']:
c.attr(rank='same')
c.node(str(n), str(n), color="#2ecc71")
for i in range(1, 1 + layers['n_layer']):
with g.subgraph(name="hidden" + str(i)) as c:
name_layer = 'hidden' + str(i)
for n in layers[name_layer]:
c.attr(rank='same')
c.node(str(n), str(n), color="#3498db")
with g.subgraph(name="output") as c:
for n in layers['output']:
c.attr(rank='same')
c.node(str(n), str(n), color="red")
for edge in layers['connections']:
g.edge(str(edge[0]), str(edge[1]), arrowhead="none")
g.render(filename, view=True)
with open('../../winner_genome', 'rb') as f:
winner = pickle.load(f)
config_file = "../../config"
config = neat.Config(MyGenome, EliteReproduction, neat.DefaultSpeciesSet,
neat.DefaultStagnation, config_file)
plot_layers(winner, config, 'prova')
def run():
# Load the config file, which is assumed to live in
# the same directory as this script.
local_dir = os.path.dirname(__file__)
config_path = os.path.join(local_dir, 'config')
config = neat.Config(LanderGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
config_path)
pop = neat.Population(config)
stats = neat.StatisticsReporter()
pop.add_reporter(stats)
pop.add_reporter(neat.StdOutReporter(True))
# Checkpoint every 25 generations or 900 seconds.
pop.add_reporter(neat.Checkpointer(25, 900))
# Run until the winner from a generation is able to solve the environment
# or the user interrupts the process.
ec = PooledErrorCompute()
while 1:
try:
gen_best = pop.run(ec.evaluate_genomes, 5)
#print(gen_best)
visualize.plot_stats(stats,
ylog=False,
view=False,
filename="fitness.svg")
plt.plot(ec.episode_score, 'g-', label='score')
plt.plot(ec.episode_length, 'b-', label='length')
plt.grid()
plt.legend(loc='best')
plt.savefig("scores.svg")
plt.close()
mfs = sum(stats.get_fitness_mean()[-5:]) / 5.0
print("Average mean fitness over last 5 generations: {0}".format(
mfs))
mfs = sum(stats.get_fitness_stat(min)[-5:]) / 5.0
print(
"Average min fitness over last 5 generations: {0}".format(mfs))
# Use the best genomes seen so far as an ensemble-ish control system.
best_genomes = stats.best_unique_genomes(3)
best_networks = []
for g in best_genomes:
best_networks.append(
neat.nn.FeedForwardNetwork.create(g, config))
solved = True
best_scores = []
for k in range(100):
observation = env.reset()
score = 0
step = 0
while 1:
step += 1
# Use the total reward estimates from all five networks to
# determine the best action given the current state.
votes = np.zeros((4, ))
for n in best_networks:
output = n.activate(observation)
votes[np.argmax(output)] += 1
best_action = np.argmax(votes)
observation, reward, done, info = env.step(best_action)
score += reward
env.render()
if done:
break
ec.episode_score.append(score)
ec.episode_length.append(step)
best_scores.append(score)
avg_score = sum(best_scores) / len(best_scores)
print(k, score, avg_score)
if avg_score < 200:
solved = False
break
if solved:
print("Solved.")
# Save the winners.
for n, g in enumerate(best_genomes):
name = 'winner-{0}'.format(n)
with open(name + '.pickle', 'wb') as f:
pickle.dump(g, f)
visualize.draw_net(config,
g,
view=False,
filename=name + "-net.gv")
visualize.draw_net(config,
g,
view=False,
filename=name + "-net-enabled.gv",
show_disabled=False)
visualize.draw_net(config,
g,
view=False,
filename=name +
"-net-enabled-pruned.gv",
show_disabled=False,
prune_unused=True)
break
except KeyboardInterrupt:
print("User break.")
break
env.close()
def run(n_generations, n_processes):
# Load the config file, which is assumed to live in
# the same directory as this script.
config_path = os.path.join(os.path.dirname(__file__), "neat.cfg")
config = neat.Config(
neat.DefaultGenome,
neat.DefaultReproduction,
neat.DefaultSpeciesSet,
neat.DefaultStagnation,
config_path,
)
envs = [
t_maze.TMazeEnv(init_reward_side=i, n_trials=100)
for i in [1, 0, 1, 0]
]
evaluator = MultiEnvEvaluator(make_net,
activate_net,
envs=envs,
batch_size=batch_size,
max_env_steps=1000)
if n_processes > 1:
pool = multiprocessing.Pool(processes=n_processes)
def eval_genomes(genomes, config):
fitnesses = pool.starmap(evaluator.eval_genome,
((genome, config)
for _, genome in genomes))
for (_, genome), fitness in zip(genomes, fitnesses):
genome.fitness = fitness
else:
def eval_genomes(genomes, config):
for i, (_, genome) in enumerate(genomes):
try:
genome.fitness = evaluator.eval_genome(genome,
config,
debug=DEBUG
and i % 100 == 0)
except Exception as e:
print(genome)
raise e
pop = neat.Population(config)
stats = neat.StatisticsReporter()
pop.add_reporter(stats)
reporter = neat.StdOutReporter(True)
pop.add_reporter(reporter)
logger = LogReporter("./logs/adaptive.json", evaluator.eval_genome)
pop.add_reporter(logger)
winner = pop.run(eval_genomes, n_generations)
print(winner)
final_performance = evaluator.eval_genome(winner, config)
print("Final performance: {}".format(final_performance))
generations = reporter.generation + 1
return generations
def run():
# Load the config file, which is assumed to live in
# the same directory as this script.
local_dir = os.path.dirname(__file__)
config_path = os.path.join(local_dir, 'config')
config = neat.Config(LanderGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
config_path)
pop = neat.Population(config)
stats = neat.StatisticsReporter()
pop.add_reporter(stats)
pop2 = neat.Population(config)
stats2 = neat.StatisticsReporter()
pop.add_reporter(stats)
pop2.add_reporter(stats2)
pop.add_reporter(neat.StdOutReporter(True))
# Checkpoint every 25 generations or 900 seconds.
rep = neat.Checkpointer(25, 900)
pop.add_reporter(rep)
# asigna un gen_best para poder cargar los demás desde syn
for g in itervalues(pop.population):
gen_best = g
g.fitness = -10000000.0
# Run until the winner from a generation is able to solve the environment
# or the user interrupts the process.
ec = PooledErrorCompute()
temp = 0
while 1:
try:
if temp >= 0:
# TODO: FUNCION DE SINCRONIZACION CON SINGULARITY
# Lee en pop2 el último checkpoint desde syn
# Hace request de getLastParam(process_hash,use_current) a syn TODO: HACER PROCESS CONFIGURABLE Y POR HASH no por id
res = requests.get(
my_url +
"/processes/1?username=harveybc&pass_hash=$2a$04$ntNHmofQoMoajG89mTEM2uSR66jKXBgRQJnCgqfNN38aq9UkN4Y6q&process_hash=ph"
)
cont = res.json()
print('\ncurrent_block_performance =',
cont['result'][0]['current_block_performance'])
print('\nlast_optimum_id =',
cont['result'][0]['last_optimum_id'])
last_optimum_id = cont['result'][0]['last_optimum_id']
# Si el perf reportado pop2_champion_fitness > pop1_champion_fitness
best_fitness = gen_best.fitness
print('\nbest_fitness =', best_fitness)
if cont['result'][0][
'current_block_performance'] > best_fitness:
# hace request GetParameter(id)
res_p = requests.get(
my_url + "/parameters/" + str(last_optimum_id) +
"?username=harveybc&pass_hash=$2a$04$ntNHmofQoMoajG89mTEM2uSR66jKXBgRQJnCgqfNN38aq9UkN4Y6q&process_hash=ph"
)
cont_param = res_p.json()
# descarga el checkpoint del link de la respuesta si cont.parameter_link
print('\ncont_param =', cont_param)
print('\nmigrations =')
if cont_param['result'][0]['parameter_link'] is not None:
genom_data = requests.get(
cont_param['result'][0]['parameter_link']).content
with open('remote_reps', 'wb') as handler:
handler.write(genom_data)
handler.close()
# carga genom descargado en nueva población pop2
with open('remote_reps', 'rb') as f:
remote_reps = pickle.load(f)
# OP.MIGRATION: Reemplaza el peor de la especie pop1 más cercana por el nuevo chmpion de pop2 como http://neo.lcc.uma.es/Articles/WRH98.pdf
# para cada elemento de remote_reps, busca el closer, si remote fitness > local, lo reemplaza
for i in range(len(remote_reps)):
closer = None
min_dist = None
for g in itervalues(pop.population):
if g not in remote_reps:
dist = g.distance(remote_reps[i],
config.genome_config)
else:
dist = 100000000
# do not count already migrated remote_reps
if closer is None or min_dist is None:
closer = deepcopy(g)
min_dist = dist
if dist < min_dist:
closer = deepcopy(g)
min_dist = dist
# For the best genom in position 0
if i == 0:
tmp_genom = deepcopy(remote_reps[i])
# Hack: overwrites original genome key with the replacing one
tmp_genom.key = closer.key
pop.population[closer.key] = deepcopy(
tmp_genom)
print(" gen_best=", closer.key)
pop.best_genome = deepcopy(tmp_genom)
gen_best = deepcopy(tmp_genom)
else:
# si el remote fitness>local, reemplazar el remote de pop2 en pop1
if closer is not None:
if closer not in remote_reps:
# ADicionada condición para que NO migre el campeón, solo los reps (prueba)
if closer.fitness is not None and remote_reps[
i].fitness is not None:
if remote_reps[
i].fitness > closer.fitness:
tmp_genom = deepcopy(
remote_reps[i])
# Hack: overwrites original genome key with the replacing one
tmp_genom.key = closer.key
pop.population[
closer.key] = deepcopy(
tmp_genom)
print("Replaced=", closer.key)
# actualiza gen_best y best_genome al remoto
pop.best_genome = deepcopy(
tmp_genom)
gen_best = deepcopy(tmp_genom)
if closer.fitness is None:
tmp_genom = deepcopy(
remote_reps[i])
# Hack: overwrites original genome key with the replacing one
tmp_genom.key = len(
pop.population) + 1
pop.population[
tmp_genom.key] = tmp_genom
print(
"Created Por closer.fitness=NONE : ",
tmp_genom.key)
# actualiza gen_best y best_genome al remoto
pop.best_genome = deepcopy(
tmp_genom)
gen_best = deepcopy(tmp_genom)
else:
#si closer está en remote_reps es porque no hay ningun otro cercano así que lo adiciona
tmp_genom = deepcopy(remote_reps[i])
# Hack: overwrites original genome key with the replacing one
tmp_genom.key = len(pop.population) + 1
pop.population[
tmp_genom.key] = tmp_genom
print(
"Created por Closer in rempte_reps=",
tmp_genom.key)
# actualiza gen_best y best_genome al remoto
pop.best_genome = deepcopy(tmp_genom)
gen_best = deepcopy(tmp_genom)
#ejecuta speciate
pop.species.speciate(config, pop.population,
pop.generation)
print("\nSpeciation after migration done")
# Si el perf reportado es menor pero no igual al de pop1
if cont['result'][0][
'current_block_performance'] < best_fitness:
# Obtiene remote_reps
# hace request GetParameter(id)
remote_reps = None
res_p = requests.get(
my_url + "/parameters/" + str(last_optimum_id) +
"?username=harveybc&pass_hash=$2a$04$ntNHmofQoMoajG89mTEM2uSR66jKXBgRQJnCgqfNN38aq9UkN4Y6q&process_hash=ph"
)
cont_param = res_p.json()
# descarga el checkpoint del link de la respuesta si cont.parameter_link
print('\ncont_param =', cont_param)
print('\nmigrations =')
if cont_param['result'][0]['parameter_link'] is not None:
genom_data = requests.get(
cont_param['result'][0]['parameter_link']).content
with open('remote_reps', 'wb') as handler:
handler.write(genom_data)
handler.close()
# carga genom descargado en nueva población pop2
with open('remote_reps', 'rb') as f:
remote_reps = pickle.load(f)
#Guarda los mejores reps
reps_local = []
reps = [gen_best]
# Para cada especie, adiciona su representative a reps
for sid, s in iteritems(pop.species.species):
#print("\ns=",s)
if s.representative not in reps_local:
reps_local.append(
pop.population[s.representative.key])
reps_local[len(reps_local) - 1] = deepcopy(
pop.population[s.representative.key])
# TODO: Conservar los mejores reps, solo reemplazarlos por los mas cercanos
if remote_reps is None:
for l in reps_local:
reps.append(l)
reps[len(reps) - 1] = deepcopy(l)
else:
# para cada reps_local l
for l in reps_local:
# busca el closer a l en reps_remote
for i in range(len(remote_reps)):
closer = None
min_dist = None
for g in reps_local:
if g not in remote_reps:
dist = g.distance(
remote_reps[i],
config.genome_config)
else:
dist = 100000000
# do not count already migrated remote_reps
if closer is None or min_dist is None:
closer = deepcopy(g)
min_dist = dist
if dist < min_dist:
closer = deepcopy(g)
min_dist = dist
# si closer is in reps
if closer in reps:
# adiciona l a reps si ya no estaba en reps
if l not in reps:
reps.append(l)
reps[len(reps) - 1] = deepcopy(l)
# sino
else:
# si l tiene más fitness que closer,
if closer.fitness is not None and l.fitness is not None:
if l.fitness > pop.population[
closer.key].fitness:
# adiciona l a reps si ya no estaba en reps
if l not in reps:
reps.append(l)
reps[len(reps) - 1] = deepcopy(l)
# sino
else:
# adiciona closer a reps si ya no estaba en reps
if l not in reps:
reps.append(
pop.population[closer.key])
reps[len(reps) - 1] = deepcopy(
pop.population[closer.key])
# Guarda checkpoint de los representatives de cada especie y lo copia a ubicación para servir vía syn.
# rep.save_checkpoint(config,pop,neat.DefaultSpeciesSet,rep.current_generation)
print("\nreps=", reps)
filename = '{0}{1}'.format("reps-", rep.current_generation)
with open(filename, 'wb') as f:
pickle.dump(reps, f)
# Hace request de CreateParam a syn
form_data = {
"process_hash":
"ph",
"app_hash":
"ah",
"parameter_link":
my_url + "/genoms/" + filename,
"parameter_text":
pop.best_genome.key,
"parameter_blob":
"",
"validation_hash":
"",
"hash":
"h",
"performance":
best_fitness,
"redir":
"1",
"username":
"******",
"pass_hash":
"$2a$04$ntNHmofQoMoajG89mTEM2uSR66jKXBgRQJnCgqfNN38aq9UkN4Y6q"
}
# TODO: COLOCAR DIRECCION CONFIGURABLE
res = requests.post(
my_url +
"/parameters?username=harveybc&pass_hash=$2a$04$ntNHmofQoMoajG89mTEM2uSR66jKXBgRQJnCgqfNN38aq9UkN4Y6q&process_hash=ph",
data=form_data)
res_json = res.json()
# TODO FIN: FUNCION DE SINCRONIZACION CON SINGULARITY
temp = temp + 1
gen_best = pop.run(ec.evaluate_genomes, 5)
#print(gen_best)
visualize.plot_stats(stats,
ylog=False,
view=False,
filename="fitness.svg")
plt.plot(ec.episode_score, 'g-', label='score')
plt.plot(ec.episode_length, 'b-', label='length')
plt.grid()
plt.legend(loc='best')
plt.savefig("scores.svg")
plt.close()
mfs = sum(stats.get_fitness_mean()[-3:]) / 3.0
print("Average mean fitness over last 3 generations: {0}".format(
mfs))
mfs = sum(stats.get_fitness_stat(min)[-3:]) / 3.0
print(
"Average min fitness over last 3 generations: {0}".format(mfs))
# Use the best genome to evaluate an environment.
best_genomes = stats.best_unique_genomes(3)
solved = True
best_scores = []
observation = env.reset()
score = 0.0
step = 0
gen_best_nn = neat.nn.FeedForwardNetwork.create(gen_best, config)
while 1:
step += 1
output = gen_best_nn.activate(nn_format(observation))
best_action = np.argmax(output)
observation, reward, done, info = env.step(best_action)
score += reward
env.render()
if done:
break
ec.episode_score.append(score)
ec.episode_length.append(step)
best_scores.append(score)
avg_score = sum(best_scores) / len(best_scores)
print("Training Set Score =", score, " avg_score=", avg_score)
#Calculate the validation set score
best_genomes = stats.best_unique_genomes(3)
solved = True
best_scores = []
observation = env_v.reset()
score = 0.0
step = 0
gen_best_nn = neat.nn.FeedForwardNetwork.create(gen_best, config)
while 1:
step += 1
output = gen_best_nn.activate(nn_format(observation))
best_action = np.argmax(output)
observation, reward, done, info = env_v.step(best_action)
score += reward
env_v.render()
if done:
break
best_scores.append(score)
avg_score = sum(best_scores) / len(best_scores)
print("Validation Set Score =", score, " avg_score=", avg_score)
if avg_score < 2000000000:
solved = False
if solved:
print("Solved.")
# Save the winners.
for n, g in enumerate(best_genomes):
name = 'winner-{0}'.format(n)
with open(name + '.pickle', 'wb') as f:
pickle.dump(g, f)
visualize.draw_net(config,
g,
view=False,
filename=name + "-net.gv")
visualize.draw_net(config,
g,
view=False,
filename=name + "-net-enabled.gv",
show_disabled=False)
visualize.draw_net(config,
g,
view=False,
filename=name +
"-net-enabled-pruned.gv",
show_disabled=False,
prune_unused=True)
break
except KeyboardInterrupt:
print("User break.")
break
env.close()
def evaluate(file,
error_rates,
error_mode,
n_games,
n_jobs,
verbose,
file_suffix='',
transfer_to_distance=None):
time_id = datetime.now()
# Load the corresponding config files
savedir = file[:file.rfind("/")]
if not os.path.exists("%s/config.json" % savedir):
raise ValueError("Configuration file does not exist (%s)." %
("%s/config.json" % savedir))
with open("%s/config.json" % savedir) as f:
config = json.load(f)
config = check_config(config)
# Load the genome to be evaluated
if not os.path.exists(file):
raise ValueError("Genome file does not exist.")
with open(file, "rb") as f:
genome = pickle.load(f)
if not os.path.exists("%s/population-config" % savedir):
raise ValueError("Population configuration file does not exist.")
population_config = neat.Config(neat.DefaultGenome,
neat.DefaultReproduction,
neat.DefaultSpeciesSet,
neat.DefaultStagnation,
"%s/population-config" % savedir)
if config["Training"]["network_type"] == 'ffnn':
if transfer_to_distance is None:
net = SimpleFeedForwardNetwork.create(genome, population_config)
code_distance = config["Physics"]["distance"]
elif transfer_to_distance > config["Physics"]["distance"]:
generate_config_file(savedir, config, transfer_to_distance)
pop_config_transferred = neat.Config(
neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation, savedir +
"/population-config-temp-d" + str(transfer_to_distance))
new_genome = pop_config_transferred.genome_type(0)
new_genome.configure_new(pop_config_transferred.genome_config)
new_genome.connections = {}
new_genome.nodes = {}
transplantate(pop_config_transferred.genome_config, new_genome,
transfer_to_distance, genome,
config["Physics"]["distance"])
net = SimpleFeedForwardNetwork.create(new_genome,
pop_config_transferred)
code_distance = transfer_to_distance
elif config["Training"]["network_type"] == 'cppn':
# HyperNEAT: possibility of evaluating a CPPN trained on d=3 data on d>3 data
if transfer_to_distance is None:
code_distance = config["Physics"]["distance"]
connection_weight_scale = 1
elif transfer_to_distance > config["Physics"]["distance"]:
code_distance = transfer_to_distance
# As they are more connections, in larger codes, we need to scale down the connection weight by this factor
connection_weight_scale = config["Physics"][
"distance"]**2 / transfer_to_distance**2
#connection_weight_scale = 0.01
else:
raise ValueError(
"Transfer knwoledge can only be done to higher distance codes."
)
if config["Training"]["substrate_type"] == 0:
substrate = SubstrateType0(
code_distance,
config["Training"]["rotation_invariant_decoder"])
elif config["Training"]["substrate_type"] == 1:
substrate = SubstrateType1(code_distance)
#print(code_distance, connection_weight_scale)
cppn_network = FeedForwardNetwork.create(genome, population_config)
net = PhenotypeNetwork.create(cppn_network, substrate,
connection_weight_scale)
# DIRTY: To ensure that samples are generated according to transfer_to_distance
config["Physics"]["distance"] = code_distance
## (PARALLEL) EVALUATION LOOP
fitness = []
results = {
"fitness": [],
"error_rate": [],
"outcome": [],
"nsteps": [],
"initial_qubits_flips": []
}
# with statement to close properly the parallel processes
with Pool(n_jobs) as pool:
# Game evaluation
for error_rate in error_rates:
fitness.append(0)
jobs = []
for i in range(n_games):
#
jobs.append(
pool.apply_async(get_fitness,
(net, config, error_rate, error_mode)))
for job in jobs:
output, errors_id = job.get(timeout=None)
fitness[-1] += output["fitness"]
for k, v in output.items():
results[k].append(v)
results["initial_qubits_flips"].append(errors_id)
fitness[-1] /= n_games
print("Evaluation on error_rate=%.2f is done, %.2f success." %
(error_rate, fitness[-1]))
elapsed = datetime.now() - time_id
print("Total running time:", elapsed.seconds, ":",
elapsed.microseconds)
# Always overwrite the result of evaluation
# Synthesis report
if transfer_to_distance is not None:
file_suffix += ".transfered_distance%i" % transfer_to_distance
savefile = "%s_evaluation.ngames=%i.errormode=%i.%s.csv" % (
file.replace(".pkl", ""), n_games, error_mode, file_suffix)
if os.path.exists(savefile):
print("Deleting evaluation file %s" % savefile)
os.remove(savefile)
print([error_rates, fitness])
df = pd.DataFrame(list(zip(error_rates, fitness)),
columns=["error_rate", "mean_fitness"])
df.to_csv(savefile)
# Detailed report
savefile = "%s_detailed_results_evaluation.ngames=%i.%s.csv" % (
file.replace(".pkl", ""), n_games, file_suffix)
if os.path.exists(savefile):
print("Deleting evaluation file %s" % savefile)
os.remove(savefile)
pd.DataFrame.from_dict(results).to_csv(savefile)
return error_rates, fitness
net = neat.nn.FeedForwardNetwork.create(genome, config)
genome.fitness = 0
for _ in range(100):
action = net.activate(observation)
observation, reward, done, info = env.step(action)
genome.fitness += reward
if done:
observation = env.reset()
break
# Load configuration.
config = neat.Config(neat.DefaultGenome, neat.MctsReproductionWeightEvolution,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
'config-feedforward')
# Create the population, which is the top-level object for a NEAT run.
p = neat.Population(config)
# Add a stdout reporter to show progress in the terminal.
p.add_reporter(neat.StdOutReporter(False))
# Run until a solution is found.
winner = p.run(eval_genomes, 200)
# Display the winning genome.
print('\nBest genome:\n{!s}'.format(winner))
# Show output of the most fit genome against training data.
def run():
# Load the config file
local_dir = os.path.dirname(__file__)
if (type == "rnn"):
config_path = os.path.join(local_dir, 'config-ctrnn')
else:
config_path = os.path.join(local_dir, 'config-feed')
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
config_path)
pop = neat.Population(config)
stats = neat.StatisticsReporter()
pop.add_reporter(stats)
pop.add_reporter(neat.StdOutReporter(True))
if 1:
winner = pop.run(eval_genomes, num_generation)
else:
pe = neat.ParallelEvaluator(4, eval_genome)
winner = pop.run(pe.evaluate, num_generation)
# Save the winner.
with open('results/winner-rnn', 'wb') as f:
pickle.dump(winner, f)
eval_genome(winner, config, True, fixed=0)
print(stats)
#plt.figure("NEAT (Population's average/std. dev and best fitness)")
#plt.rcParams["figure.figsize"] = [6,4]
plt.rcParams.update({'font.size': 15})
plt.figure(
"NEAT fitnesses") # (Population's average/std. dev and best fitness)")
avg_fitnesses = {
"total": [],
"rotation": [],
"energy": [],
"time": [],
"accuracy": []
}
for i in range(0, num_generation):
avg_fitnesses["total"].append(
np.mean(fitnesses["total"][i * pop_size:i * pop_size + pop_size -
1]))
avg_fitnesses["time"].append(
np.mean(fitnesses["time"][i * pop_size:i * pop_size + pop_size -
1]))
avg_fitnesses["rotation"].append(
np.mean(fitnesses["rotation"][i * pop_size:i * pop_size +
pop_size - 1]))
avg_fitnesses["energy"].append(
np.mean(fitnesses["energy"][i * pop_size:i * pop_size + pop_size -
1]))
avg_fitnesses["accuracy"].append(
np.mean(fitnesses["accuracy"][i * pop_size:i * pop_size +
pop_size - 1]))
#print(avg_fitnesses["accuracy"])
#print(avg_fitnesses)
#plt.plot([i for i in range(0, len(avg_fitnesses["total"]))], avg_fitnesses["total"], 'b-', label="fitness")
plt.plot([i for i in range(0, len(avg_fitnesses["energy"]))],
avg_fitnesses["energy"],
color="#000000",
label="energy",
linestyle='-')
#plt.plot(generation, avg_fitness - stdev_fitness, 'g-.', label="-1 sd")
#plt.plot(generation, avg_fitness + stdev_fitness, 'g-.', label="+1 sd")
#plt.plot(generation, best_fitness, 'r-', label="best")
#plt.title("Population's average/std. dev and best fitness")
plt.xlabel("Generations")
plt.ylabel("Fitness", labelpad=0)
plt.grid()
plt.legend(loc="best")
plt.savefig("energy_" + type_optimization + ".png")
view = True
if view:
plt.show()
#plt.close()
fig = None
plt.plot([i for i in range(0, len(avg_fitnesses["accuracy"]))],
avg_fitnesses["accuracy"],
color="#808080",
linestyle="-",
label="accuracy",
lw=0.9)
plt.plot([i for i in range(0, len(avg_fitnesses["time"]))],
avg_fitnesses["time"],
color="#D3D3D3",
linestyle="-",
label="time",
lw=0.9)
plt.plot([i for i in range(0, len(avg_fitnesses["rotation"]))],
avg_fitnesses["rotation"],
color="#000000",
linestyle="-",
label="rotation",
lw=0.9)
plt.xlabel("Generations")
plt.ylabel("Fitness")
plt.grid()
plt.legend(loc="best")
plt.savefig("accuracy_time_rotation_" + type_optimization + ".png")
view = True
if view:
plt.show()
fig = None
if (1):
visualize.plot_stats(stats,
ylog=True,
view=False,
filename="results/rnn-fitness.svg")
visualize.plot_species(stats,
view=False,
filename="results/rnn-speciation.svg")
node_names = {
0: 'omega',
1: 'theta',
2: 'psi',
3: 'fi',
4: 'ro',
5: 'epsilon',
-3: 'x',
-2: 'y',
-1: 'z'
}
visualize.draw_net(config,
winner,
view=False,
node_names=node_names,
filename="results/winner-rnn.gv")
visualize.draw_net(config,
winner,
view=False,
node_names=node_names,
filename="results/winner-rnn-enabled.gv",
show_disabled=False)
visualize.draw_net(config,
winner,
view=False,
node_names=node_names,
filename="results/winner-rnn-enabled-pruned.gv",
show_disabled=False,
prune_unused=True)
xor_outputs = [(0.0, ), (1.0, ), (1.0, ), (0.0, )]
def eval_genomes(genomes, config):
for genome_id, genome in genomes:
genome.fitness = 4.0
net = neat.nn.FeedForwardNetwork.create(genome, config)
for xi, xo in zip(xor_inputs, xor_outputs):
output = net.activate(xi)
# this checks the squared error between the output and input
genome.fitness -= (output[0] - xo[0])**2
# Load configuration.
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
'./temp.conf')
# Create the population, which is the top-level object for a NEAT run.
p = neat.Population(config)
# Add a stdout reporter to show progress in the terminal.
p.add_reporter(neat.StdOutReporter(False))
# Run until a solution is found.
winner = p.run(eval_genomes, n=6)
# Display the winning genome.
print('\nBest genome:\n{!s}'.format(winner))
# Show output of the most fit genome against training data.
def __init__(self, agent_type, name, ea_framework, idx):
self.nets = []
self.genomes = []
self.genome_ids = []
self.configs = []
self.fitnesses = []
self.prev_nets = []
self.prev_genomes = []
self.prev_genome_ids = []
self.prev_configs = []
self.prev_fitnesses = []
self.best_nets = []
self.ea_framework = ea_framework
self.evolved = False
self.human_left = 0.0
self.human_right = 0.0
self.idx = idx
self.type = agent_type # type = 0, type = 1
# 兩個evolution_event 是由ea_framework決定
self.evolution_event = threading.Event()
self.evolution_end_event = threading.Event()
self.evolution_event.daemon = True
self.evolution_end_event.daemon = True
self.name = name
self.neat = None
self.orientation = None
self.position = None
local_dir = os.path.dirname(__file__)
if agent_type == 0:
config_path = os.path.join(local_dir, 'config_predator')
else:
config_path = os.path.join(local_dir, 'config_prey')
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
config_path)
if not LOAD_WEIGHT:
p = neat.Population(config)
self.neat = p
else:
p = neat.Checkpointer.restore_checkpoint('/home/jaqq/catkin_ws/src/robobo/robobo_gazebo/scripts/data/' + DATA + '/checkpoints/' + self.name + '/neat-checkpoint-' + str(LOADED_GENERATION))
#p = neat.Checkpointer.restore_checkpoint('/home/jaqq/Documents/robobo_bak/scripts_without_random/checkpoints/' + self.name + '/neat-checkpoint-' + str(LOADED_GENERATION))
self.neat = p
pkl_file = open('/home/jaqq/catkin_ws/src/robobo/robobo_gazebo/scripts/data/' + DATA + '/generations/' + self.name + "/" + str(self.neat.generation - 1) + '.pkl', 'rb')
#pkl_file = open('/home/jaqq/Documents/robobo_bak/scripts_without_random/generations/' + self.name + "/" + str(self.neat.generation - 1) + '.pkl', 'rb')
data_dict = pickle.load(pkl_file)
self.nets = data_dict['nets']
self.genomes = data_dict['genomes']
self.genome_ids = data_dict['genome_ids']
self.configs = data_dict['configs']
self.fitnesses = data_dict['fitnesses']
self.best_nets = data_dict['best_nets']
if not EVALUATION:
self.neat.add_reporter(neat.StdOutReporter(True))
stats = neat.StatisticsReporter()
self.neat.add_reporter(stats)
save_dir = "checkpoints/" + self.name
save_prefix = save_dir + "/neat-checkpoint-"
self.neat.add_reporter(neat.Checkpointer(1, filename_prefix=save_prefix))
if not os.path.exists(save_dir):
os.mkdir(save_dir)
self.fitness_dir = "fitness/" + self.name
self.fitness_path = self.fitness_dir + "/fitness.txt"
if not os.path.exists(self.fitness_dir):
os.mkdir(self.fitness_dir)
if os.path.exists(self.fitness_path):
os.remove(self.fitness_path)
res.write('{0:3.3f},{1:3.3f},{2:3.3f},{3},{4:3.6f}\n'.format(best_every_generation[best_id][0],
best_every_generation[best_id][1],
best_every_generation[best_id][2],
best_every_generation[best_id][3],
best_every_generation[best_id][4]))
res.close()
#del the net not in current population
#net_id = set(net_dict.keys())
#net_to_del = net_id - genomes_id_set
#for genome_id in net_to_del:
# del net_dict[genome_id]
# Load configuration.
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
'config-mnist')
if torch.cuda.is_available():
gpu = True
print("Running on GPU!")
else:
gpu = False
print("Running on CPU!")
# reset result file
res = open("result.csv", "w")
info = open("info.txt", "w")
best = open("best.txt", "w")
res.close()
info.close()
import visualize, neat, gym, sys
if len(sys.argv) < 2:
print("Usage: {} path-to-dir".format(sys.argv[0]))
exit()
dire = sys.argv[1]
print("Dir: {}".format(dire))
dire = './' + dire + '/'
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
dire + "fc.config")
winner = None
def get_winner(p):
max_fitness = -9999999999
best_genome = None
for v in p.population:
genome = p.population[v]
if genome.fitness > max_fitness:
max_fitness = genome.fitness
winner = genome
print(max_fitness)
return ([max_fitness, winner])
mf = -999999999
def eval_genomes(genomes, config):
for genome_id, genome in genomes:
net = neat.nn.FeedForwardNetwork.create(genome, config)
state = env.reset()
total_reward = 0
done = False
while not done:
action = np.argmax(net.activate(state))
state, reward, done, _ = env.step(action)
total_reward += reward
genome.fitness = total_reward
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation, 'GymNEAT')
p = neat.Population(config)
p.add_reporter(neat.StdOutReporter(False))
winner = p.run(eval_genomes, 1000)
net = neat.nn.FeedForwardNetwork.create(winner, config)
state = env.reset()
total_reward = 0
done = False
while not done:
env.render()
action = np.argmax(net.activate(state))
state, reward, done, _ = env.step(action)
total_reward += reward
print('Total Reward:', total_reward)
def main():
#Create gym-retro environment for game of choice: DRMBM
env = retro.make(game = 'DrRobotniksMeanBeanMachine-Genesis')
imgarray = []
def eval_genomes(genomes, config):
for genome_id, genome in genomes:
# Set observation variable (Age)
ob = env.reset()
# Create a random action (generic)
ac = env.action_space.sample()
# Get the x, y and colors of the input space (from the emulator)
inx, iny, inc = env.observation_space.shape
# Divide input by 8
inx, iny, inc = int(inx/8), int(iny/8), int(inc/8)
#Create NEAT net
net = neat.nn.recurrent.RecurrentNetwork.create(genome, config)
current_max_fitness = 0
fitness_current = 0
frame = 0
counter = 0
xpos = 0
xpos_max = 0
done = False
while not done:
env.render()
frame+= 1
ob = cv2.resize(ob, (inx, iny))
ob = cv2.cvtColor(ob, cv2.COLOR_BGR2GRAY)
ob = np.reshape(ob, (inx,iny))
#cut image that NN sees in half
ob= ob[0:ob.shape[0], 0:int(ob.shape[1]/2)]
imgarray = np.ndarray.flatten(ob)
nnOutput = net.activate(imgarray)
#increment emulator by 1 step
ob, rew, done, info = env.step(nnOutput)
fitness_current += rew
if fitness_current > current_max_fitness:
current_max_fitness = fitness_current
counter = 0
else:
counter += 1
if done or counter == 250:
done = True
print(genome_id, fitness_current)
genome.fitness = fitness_current
#Config File necessary to create a NEAT
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
'config-feedforward.txt')
#Create starting population
p = neat.Population(config)
#Compute statistic for the game
p.add_reporter(neat.StdOutReporter(True))
stats = neat.StatisticsReporter()
p.add_reporter(stats)
p.add_reporter(neat.Checkpointer(10))
winner = p.run(eval_genomes)
with open('winner.pkl', 'wb') as output:
pickle.dump(winner, output, 1)
def run(file_number):
# Load the config file, which is assumed to live in
# the same directory as this script.
global clearinfo
global stage_list
global actorconfig_path
global actorconfig
stage_list = []
stage_list.append(int(file_number))
print(stage_list)
local_dir = os.path.dirname(__file__)
actorconfig_path = os.path.join(local_dir, 'config-feed-action')
actorconfig = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
actorconfig_path)
config_path = os.path.join(local_dir, 'config-action-multi')
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
config_path)
with open('Network/noconnect', 'rb') as f:
c = pickle.load(f)
NoNets.append(c)
for i in range(15):
index = i + 1
with open('Network/winner-enemyNo' + str(index), 'rb') as f:
c = pickle.load(f)
EnemyNets.append(c)
print('Enemy Nets stanby complete.')
sleep(5)
for i in range(15):
index = i + 1
with open('Network/winner-jumpNo' + str(index), 'rb') as f:
c = pickle.load(f)
JumpNets.append(c)
print('Jump Nets stanby complete.')
pop = neat.Population(config)
stats = neat.StatisticsReporter()
pop.add_reporter(stats)
pop.add_reporter(neat.StdOutReporter(True))
winner = pop.run(eval_genomes, 200)
if (trainend == False):
clearinfo.append((-1, round(nowBest, 3)))
with open('Results/multisingles' + str(file_number) + '.txt',
'a') as f:
f.write("(-1," + str(round(nowBest, 3)) + ")")
f.write('\n')
# Save the winner.
#with open('Network/winner-multi-No'+file_number, 'wb') as f:
# pickle.dump(winner, f)
#print(winner)
print("Now Multi Stage with Single Stage Networks version Clear info is ")
print(clearinfo)
#visualize.plot_stats(stats, ylog=False, view=False, filename="feed-fitness-fs-sigmoid-50hidden-50.svg")
#visualize.plot_species(stats, view=False, filename="feed-speciation-fs-sigmoid-50hidden-150.svg")
node_names = {-1: '(-3,-3)', -2: '(-2,-3)', -3: '(-1,-3)', -4: '(0,-3)', -5: '(1,-3)',-6: '(2,-3)', -7: '(3,-3)',\
-8: '(-3,-2)', -9: '(-2,-2)', -10: '(-1,-2)', -11: '(0,-2)', -12: '(1,-2)', -13:'(2,-2)', -14:'(3,-2)',\
-15: '(-3,-1)', -16: '(-2,-1)', -17:'(-1,-1)',-18:'(0,-1)',-19:'(1,-1)',-20:'(2,-1)',-21:'(3,-1)',\
-22: '(-3,0)', -23: '(-2,0)', -24:'(-1,0)',-25:'(0,0)',-26:'(1,0)',-27:'(2,0)',-28:'(3,0)',\
-29: '(-3,1)', -30: '(-2,1)', -31:'(-1,1)',-32:'(0,1)',-33:'(1,1)',-34:'(2,1)',-35:'(3,1)',\
-36: '(-3,2)', -37: '(-2,2)', -38:'(-1,2)',-39:'(0,2)',-40:'(1,2)',-41:'(2,2)',-42:'(3,2)',\
-43: '(-3,3)', -44: '(-2,3)', -45:'(-1,3)',-46:'(0,3)',-47:'(1,3)',-48:'(2,3)',-49:'(3,3)',\
-50:'vx',-51: 'vy',-52:'now_mode',-53:'on_floor?',-54:'now_dive?',-55:'now_HP',-56:'can_rejump?',-57:'can_attack?',\
0: 'None',1:'←',2:'↑',3:'→',4:'↓',5:'↑+←',6:'↑+→',7:'↓+→',8:'↓+←',9:'Z',10:'←+Z',11:'↑+Z',\
12:'→+Z',13:'↓+Z',14:'↑+←+Z',15:'↑+→+Z',16:'↓+→+Z',17:'↓+←+Z',18:'C',19:'←+C',20:'→+C'
}
def run(config_file, rep_type='Tournament', restore_checkpoint=False):
last_best_pso_cost = 0
# Load configuration.
if rep_type == 'Tournament':
rep_class = TournamentReproduction
else:
rep_class = neat.DefaultReproduction
gen_class = neat.DefaultGenome
pop_class = TimePopulation
config = neat.Config(gen_class, rep_class, neat.DefaultSpeciesSet,
neat.DefaultStagnation, config_file)
# Create the population, which is the top-level object for a NEAT run.
if restore_checkpoint:
p = neat.Checkpointer.restore_checkpoint('neat-checkpoint-35')
else:
p = pop_class(config)
p.allow_regeneration(False)
name_run = "output.txt"
# Add a stdout reporter to show progress in the terminal.
p.add_reporter(FileReporter(True, name_run))
p.add_reporter(neat.StdOutReporter(True))
stats = neat.StatisticsReporter()
p.add_reporter(stats)
p.add_reporter(MyCheckpointer(checkpoint_interval=50))
pe = neat.ParallelEvaluator(n_workers, evaluator.eval_genome)
winner = None
for i in range(n_max_gen):
winner = p.run(pe.evaluate, step_neat_gen)
if last_best_pso_cost < winner.fitness:
with open('last_winner', 'wb') as f:
pickle.dump(winner, f)
best_genome_weight = []
genome_key_set = []
for key_id in winner.connections.keys():
genome_key_set.append(key_id)
best_genome_weight.append(winner.connections[key_id].weight)
fitness = fitness_manager
dimension = len(best_genome_weight)
bounds = get_bounds(dimension)
options = {'c1': 0.5, 'c2': 0.3, 'w': 0.9, 'k': 2, 'p': 2}
optimizer = MyGlobalBestPSO(n_particles=step_pso_pop,
dimensions=dimension,
options=options,
bounds=bounds)
optimizer.set_reporter_name(name_run)
optimizer.swarm.position[0] = best_genome_weight
cost, pos = optimizer.optimize(fitness, iters=step_pso_gen)
last_best_pso_cost = -cost
if -cost > p.best_genome.fitness:
p.population[p.best_genome.key] = insert_weights(
pos, p.population[p.best_genome.key])
# Save the winner
with open('winner_genome', 'wb') as f:
pickle.dump(winner, f)
gym.register(id='tiling-pattern11x11-block-sum-v0',
entry_point='gym_multi_robot.envs:WeightedSumTilingPatternEnv',
kwargs={
'env_storage_path': 'tiles11x11_block.pickle',
'game_cls': DifferenceTilingPatternGame
})
if __name__ == '__main__':
""" Experiment where fitness is gotten by summing over all fitnesses at every timestep, multiplied with time step
to give an higher importance to later grids.
"""
env = gym.make('tiling-pattern11x11-block-sum-v0')
runner = NEATSwarmExperimentRunner(env, num_steps)
# Create learning configuration.
local_dir = os.path.dirname(__file__)
config_path = os.path.join(local_dir, config_name)
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
config_path)
config.fitness_threshold = 3000 * 100 * 3000
# Create and run experiment.
experiment = SingleExperiment(config, runner, num_generations,
experiment_name)
for i in range(num_runs):
experiment.run(experiment_name + str(i))
def call_simulate(self, data):
if data.data[0] == '-':
return
local_dir = os.path.dirname(__file__)
config_file = os.path.join(local_dir, 'config-feedforward')
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
config_file)
gene_dir = os.getenv("WALKYTO_PATH")
gene_id = self.string_decoder(data)
if gene_id == -1 or gene_id == '-1':
return
# os.write(sys.__stderr__.fileno(), "%s\n"%gene_id)
gene_f = open('%s/src/genes/%s' % (gene_dir, gene_id),'rb')
genome = pickle.load(gene_f)
gene_f.close()
net = neat.nn.FeedForwardNetwork.create(genome, config)
#nn.recurrent.RecurrentNetwork ************** param3
#-------------------------------------------------------------------------------------------------
now = rospy.Time.now()
duration = rospy.Duration(180)
then = now + duration
#self.gazebo_init()
pos_init = self.get_pose()
#--------------------------------------------------------------------------------------------------
dur = rospy.Duration(0.3); gap = rospy.Duration(0)
MAX_TORQUE = 3.0
while(then > now):
joint_efforts = net.activate(self.joint_states)
for i in range(10):
if joint_efforts[i] > MAX_TORQUE:
joint_efforts[i] = MAX_TORQUE
self.efforts_caller(joint_efforts, dur)#0.05sec
rospy.sleep(dur)
# print "input:", self.joint_states
# print "output:", joint_efforts
now = rospy.Time.now()
if gap == rospy.Duration(0):
if then-now > duration + rospy.Duration(5):
gap = then-now
elif gap > rospy.Duration(0):
now = now + gap - duration
# print self.dup_num, gap.to_sec(), then.to_sec(), now.to_sec()
#--------------------------------------------------------------------------------------------------
pos_end = self.get_pose()
dist = (pos_end.x - pos_init.x)
self.fitness = dist
def get_config():
local_dir = os.path.dirname(__file__)
path = os.path.join(local_dir, config_path)
return neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
config_path)
def run(config_file):
# Load configuration.
config = neat.Config(
neat.DefaultGenome,
neat.DefaultReproduction,
neat.DefaultSpeciesSet,
neat.DefaultStagnation,
config_file,
)
# Create the population, which is the top-level object for a NEAT run.
p = neat.Population(config)
# Add a stdout reporter to show progress in the terminal.
p.add_reporter(neat.StdOutReporter(True))
stats = neat.StatisticsReporter()
p.add_reporter(stats)
# Run for up to 300 generations.
pe = neat.ParallelEvaluator(multiprocessing.cpu_count(), eval_genome)
winner = p.run(pe.evaluate, 2000)
# Display the winning genome.
print("\nBest genome:\n{!s}".format(winner))
# Show output of the most fit genome against training data.
print("\nOutput:")
winner_net = neat.nn.FeedForwardNetwork.create(winner, config)
for xi, xo in zip(xor_inputs, xor_outputs):
output = winner_net.activate(xi)
print("input {!r}, expected output {!r}, got {!r}".format(
xi, xo, output))
df = pd.read_csv('csvs/' + inputs + 'lynx_test_X.csv',
sep=' ',
header=None)
new_test_X = df.values
X_test_inputs = []
for i in range(len(new_test_X)):
X_test_inputs.append(tuple(new_test_X[i]))
predictions_enn = []
for xi in X_test_inputs:
output = winner_net.activate(xi)
predictions_enn.append(output)
np.savetxt('csvs/' + inputs + 'predictions_enn.csv',
np.array(predictions_enn),
delimiter=',')
real_y = pd.read_csv('csvs/' + inputs + 'lynx_test_Y.csv',
sep=' ',
header=None)
mse = np.sum(
(np.array(real_y) - predictions_enn)**2) / (len(predictions_enn))
mae = np.average(np.abs(np.array(real_y) - predictions_enn))
print("MSE:", mse)
print("MAE:", mae)
node_names = {
-1: "t-7",
-2: "t-6",
-3: "t-5",
-4: "t-4",
-5: "t-3",
-6: "t-2",
-7: "t-1",
0: "Target"
}
visualize.draw_net(config, winner, True, node_names=node_names, fmt="png")
visualize.plot_stats(stats, ylog=False, view=False)
visualize.plot_species(stats, view=False)