def launch(args, defaults, description):
"""
Execute a complete training run.
"""
logging.basicConfig(level=logging.INFO)
parameters = process_args(args, defaults, description)
if parameters.rom.endswith('.bin'):
rom = parameters.rom
else:
rom = "%s.bin" % parameters.rom
full_rom_path = os.path.join(defaults.BASE_ROM_PATH, rom)
if parameters.deterministic:
rng = np.random.RandomState(123456)
else:
rng = np.random.RandomState()
ale = ale_python_interface.ALEInterface()
ale.setInt('random_seed', rng.randint(1000))
if parameters.display_screen:
import sys
if sys.platform == 'darwin':
import pygame
pygame.init()
ale.setBool('sound', False) # Sound doesn't work on OSX
ale.setBool('display_screen', parameters.display_screen)
ale.setFloat('repeat_action_probability',
parameters.repeat_action_probability)
ale.loadROM(full_rom_path)
num_actions = len(ale.getMinimalActionSet())
# Load VAE file
with open(parameters.vae_aux_file, 'r') as f:
aux_data = pickle.load(f)
f.close()
params = aux_data['params']
with tf.device(None): #"/gpu:0"):
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
# sess = tf.Session(config=config)
# config = tf.ConfigProto(
# device_count={'GPU': 0}
# )
sess = tf.Session(config=config)
VAE = vae.vae(params)
VAE._create_network_()
try:
sess.run(tf.global_variables_initializer())
except AttributeError:
sess.run(tf.initialize_all_variables())
saver = tf.train.Saver()
chkpt = tf.train.get_checkpoint_state(parameters.vae_file)
if chkpt and chkpt.model_checkpoint_path:
saver.restore(sess, chkpt.model_checkpoint_path)
else:
print 'No checkpoint found'
import theano
import ale_experiment
import ale_agent
import q_network
if parameters.cudnn_deterministic:
theano.config.dnn.conv.algo_bwd = 'deterministic'
if parameters.nn_file is None:
network = q_network.DeepQLearner(
VAE.z_size, 1, num_actions, parameters.phi_length,
parameters.discount, parameters.learning_rate,
parameters.rms_decay, parameters.rms_epsilon, parameters.momentum,
parameters.clip_delta, parameters.freeze_interval,
parameters.batch_size, parameters.network_type,
parameters.update_rule, parameters.batch_accumulator, rng)
else:
handle = open(parameters.nn_file, 'r')
network = cPickle.load(handle)
agent = ale_agent.NeuralAgent(
network, parameters.epsilon_start, parameters.epsilon_min,
parameters.epsilon_decay, parameters.replay_memory_size,
parameters.experiment_prefix, parameters.replay_start_size,
parameters.update_frequency, rng)
experiment = ale_experiment.ALEExperiment(
ale, agent, VAE.X_size[1], VAE.X_size[0], parameters.resize_method,
parameters.epochs, parameters.steps_per_epoch,
parameters.steps_per_test, parameters.frame_skip,
parameters.death_ends_episode, parameters.max_start_nullops, rng, VAE,
sess)
experiment.run()
tr_iters = commandline_params['tr_iters']
lr_rate = commandline_params['learning_rate']
dump_path = commandline_params['dump_path']
params = {}
params['batch_size'] = commandline_params['batch_size']
params['X_size'] = [210, 160, 3]
params['z_size'] = 30
params['beta'] = commandline_params['beta']
print('---Params used---')
print(params)
params_generated = params
VAE = vae.vae(params)
VAE._create_network_()
train_step = tf.train.AdamOptimizer(lr_rate).minimize(VAE.total_loss)
try:
sess.run(tf.global_variables_initializer())
except AttributeError:
sess.run(tf.initialize_all_variables())
saver = tf.train.Saver()
aux_data = {'params': params, 'commandline_params': commandline_params, \
'perm_train': perm_train, 'perm_valid': perm_valid, \
'magic_seed_number': magic_seed_number}
def launch(args, defaults, description):
"""
Execute a complete training run.
"""
logging.basicConfig(level=logging.INFO)
parameters = process_args(args, defaults, description)
if parameters.rom.endswith('.bin'):
rom = parameters.rom
else:
rom = "%s.bin" % parameters.rom
full_rom_path = os.path.join(defaults.BASE_ROM_PATH, rom)
if parameters.deterministic:
rng = np.random.RandomState(123456)
else:
rng = np.random.RandomState()
ale = ale_python_interface.ALEInterface()
ale.setInt('random_seed', rng.randint(1000))
if parameters.display_screen:
import sys
if sys.platform == 'darwin':
import pygame
pygame.init()
ale.setBool('sound', False) # Sound doesn't work on OSX
ale.setBool('display_screen', parameters.display_screen)
ale.setFloat('repeat_action_probability',
parameters.repeat_action_probability)
ale.loadROM(full_rom_path)
num_actions = len(ale.getMinimalActionSet())
# Load VAE file
with open(parameters.vae_aux_file, 'r') as f:
aux_data = pickle.load(f)
f.close()
params = aux_data['params']
with tf.device(None): #"/gpu:0"):
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
# sess = tf.Session(config=config)
# )
sess = tf.Session(config=config)
VAE = vae.vae(params)
VAE._create_network_()
try:
sess.run(tf.global_variables_initializer())
except AttributeError:
sess.run(tf.initialize_all_variables())
saver = tf.train.Saver()
import theano
import ale_experiment
import ale_agent
import q_network
if parameters.cudnn_deterministic:
theano.config.dnn.conv.algo_bwd = 'deterministic'
if parameters.nn_file is None:
network = q_network.DeepQLearner(
defaults.RESIZED_WIDTH, defaults.RESIZED_HEIGHT, num_actions,
parameters.phi_length, parameters.discount,
parameters.learning_rate, parameters.rms_decay,
parameters.rms_epsilon, parameters.momentum, parameters.clip_delta,
parameters.freeze_interval, parameters.batch_size,
parameters.network_type, parameters.update_rule,
parameters.batch_accumulator, rng)
else:
handle = open(parameters.nn_file, 'r')
network = cPickle.load(handle)
agent = ale_agent.NeuralAgent(
network, parameters.epsilon_start, parameters.epsilon_min,
parameters.epsilon_decay, parameters.replay_memory_size,
parameters.experiment_prefix, parameters.replay_start_size,
parameters.update_frequency, rng)
experiment = ale_experiment.ALEExperiment(
ale, agent, defaults.RESIZED_WIDTH, defaults.RESIZED_WIDTH,
parameters.resize_method, parameters.epochs,
parameters.steps_per_epoch, parameters.steps_per_test,
parameters.frame_skip, parameters.death_ends_episode,
parameters.max_start_nullops, rng, VAE, sess, defaults.VAE_REQ_STEPS,
defaults.VAE_STORAGE_SIZE)
time_str = time.strftime("%m-%d-%H-%M", time.gmtime())
vae_save_path = '%s/%s_beta%f_z%d' % (defaults.VAE_OUT_PREFIX,
rom.split('.')[0], params['beta'],
params['z_size'])
os.system('mkdir -p %s' % (vae_save_path))
experiment.run()
ckpt_path = '%s/%s_%s' % (vae_save_path, rom.split('.')[0], time_str)
print ckpt_path
if not os.path.exists(ckpt_path):
os.makedirs(ckpt_path)
saver.save(sess, '%s/checkpoint.ckpt' % (ckpt_path))