def run_experiments(output_dir):
POPULATION_SIZE = 100
NUM_GENERATIONS = 100
CROSSOVER_PROB = 0.5
MUTATION_PROBS = [0.05, 0.10, 0.20, 0.30, 0.40, 0.50]
for mutation_prob in MUTATION_PROBS:
pop, log, hof, history, best_per_gen = run(num_gen=NUM_GENERATIONS,
n=POPULATION_SIZE,
cxpb=CROSSOVER_PROB,
mutpb=mutation_prob)
best = np.asarray(hof)
gen = log.select("gen")
fitness_maxs = log.select("max")
fitness_avgs = log.select("avg")
plot_results(filename='{}train_cnn_ga_mutpb_{}.png'.
format(output_dir,
str(mutation_prob).replace('.', '_')),
gen=gen,
fitness_maxs=fitness_maxs,
fitness_avgs=fitness_avgs)
np.savetxt('{}train_cnn_ga_mutpb_{}.out'.
format(output_dir,
str(mutation_prob).replace('.', '_')), best)
# Plot the feature vectors produced by the best individual from each
# generation
for gen in range(len(best_per_gen)):
update_model_weights(model, np.asarray(best_per_gen[gen]))
feature_vectors = calculate_cnn_output(model, images)
plot_feature_vectors(feature_vectors, filename='{}feature_vectors_{}__{}.png'.\
format(output_dir, str(mutation_prob).replace('.', '_'), gen))
def ga_fitness(individual):
sessionid = multiprocessing.current_process()._identity[0]
# Instantiate the control policy
# Load the ConvNet weights
pretrained_weights_cnn = \
np.loadtxt('{}'.format(PRETRAINED_WEIGHTS_CNN))
POLICY = CNNRNNPolicy(cnn_weights=pretrained_weights_cnn)
# Instantiate the simulator environment
TORCS = pyclient.TORCS(episode_length=500,
policy=POLICY,
sessionid=sessionid)
# The GA will update the RNN parameters to evaluate candidate solutions
update_model_weights(POLICY.rnn, np.asarray(individual))
POLICY.rnn.reset_states()
# Run an episode on the track and its mirror image and measure the average
# fitness
fitness = 0
fitness1 = TORCS.run(track_id=1)
fitness2 = TORCS.run(track_id=2)
fitness = min(fitness1, fitness2)
return fitness,
def __init__(self, cnn_weights=None):
# Create the ConvNet and load the pretrained weights if provided
self.cnn = create_cnn()
if cnn_weights is not None:
update_model_weights(self.cnn, cnn_weights)
# Create the RNN
self.rnn = create_rnn()
def ga_fitness(individual):
sessionid = multiprocessing.current_process()._identity[0]
# Instantiate the control policy
# Load the ConvNet weights
pretrained_weights_cnn = \
np.loadtxt('{}'.format(PRETRAINED_WEIGHTS_CNN))
POLICY = CNNRNNPolicy(cnn_weights=pretrained_weights_cnn)
# Instantiate the simulator environment
TORCS = pyclient.TORCS(episode_length=500,
policy=POLICY,
sessionid=sessionid)
# The GA will update the RNN parameters to evaluate candidate solutions
update_model_weights(POLICY.rnn, np.asarray(individual))
POLICY.rnn.reset_states()
# Run an episode on the track and its mirror image and measure the average
# fitness
fitness = 0
fitness1 = TORCS.run(track_id=1)
fitness2 = TORCS.run(track_id=2)
fitness = min(fitness1, fitness2)
return fitness,
def ga_fitness(individual):
# Update the ConvNet parameters
update_model_weights(model, np.asarray(individual))
# Calculate the output feature vectors
feature_vectors = calculate_cnn_output(model, images)
# Check their fitness
fitness = calculate_fitness(feature_vectors)
return fitness,
import numpy as np
weights_index = -3
# Specify the path to the pretrained convolutional neural network weights here:
pretrained_weights_cnn = \
np.loadtxt('experiments/train_cnn_ga_11/train_cnn_ga_mutpb_0_2.out')
saved_weights_rnn_multiple = np.loadtxt('history.out')
saved_weights_rnn_best = np.loadtxt('weights.out')
POLICY = CNNRNNPolicy(cnn_weights=pretrained_weights_cnn)
weights = saved_weights_rnn_multiple[weights_index]
update_model_weights(POLICY.rnn, weights)
# Instantiate the simulator environment
TORCS = pyclient.TORCS(episode_length=500, policy=POLICY, sessionid=1)
# Run an episode on the track and its mirror image and measure the average
# fitness
fitness1 = TORCS.run(track_id=1)
fitness2 = TORCS.run(track_id=2)
fitness = min(fitness1, fitness2)
print('fitness 1: {}, fitness 2: {}, min fitness: {}'.format(
fitness1, fitness2, fitness))
import IPython
weights_index = -3
# Specify the path to the pretrained convolutional neural network weights here:
pretrained_weights_cnn = \
np.loadtxt('experiments/train_cnn_ga_11/train_cnn_ga_mutpb_0_2.out')
saved_weights_rnn_multiple = np.loadtxt('history.out')
saved_weights_rnn_best = np.loadtxt('weights.out')
POLICY = CNNRNNPolicy(cnn_weights=pretrained_weights_cnn)
weights = saved_weights_rnn_multiple[weights_index]
update_model_weights(POLICY.rnn, weights)
# Instantiate the simulator environment
TORCS = pyclient.TORCS(episode_length=500,
policy=POLICY,
sessionid=1)
# Run an episode on the track and its mirror image and measure the average
# fitness
fitness1 = TORCS.run(track_id=1)
fitness2 = TORCS.run(track_id=2)
fitness = min(fitness1, fitness2)
print('fitness 1: {}, fitness 2: {}, min fitness: {}'.format(
fitness1, fitness2, fitness))