def main(args):
if args.resume:
conf_file = os.path.join(args.debugging_folder, 'args.json')
assert os.path.exists(
conf_file
), "Could not find an args.json file in the debugging folder"
for k, v in logger_utils.load_args(args.debugging_folder).items():
setattr(args, k, v)
logger.debug('Configuration: {}'.format(args))
logger_utils.save_args(args, args.debugging_folder)
if 'gpu' in args.device:
agent_gpu = str(misc_utils.pick_gpu_lowest_memory())
os.environ["CUDA_VISIBLE_DEVICES"] = agent_gpu
logger.debug('Agent will be run on device /gpu:{}'.format(agent_gpu))
args.random_seed = 3 # random_seed
env_creator = environment_creator.EnvironmentCreator(args)
args.num_actions = env_creator.num_actions
args.state_shape = env_creator.state_shape
import numpy as np
# Create a set of arrays (as many as emulators) to exchange states, rewards, etc. between the agent and the emulator
n_emulators = args.n_emulator_runners * args.n_emulators_per_emulator_runner
variables = {
"s": np.zeros((n_emulators, ) + args.state_shape, dtype=np.float32),
"a": np.zeros((n_emulators), dtype=np.int32), # Actions
"r": np.zeros((n_emulators), dtype=np.float32), # Rewards
"done": np.zeros((n_emulators), dtype=np.bool)
} # Dones
sim_coordinator = SimulatorsCoordinator(
env_creator, args.n_emulators_per_emulator_runner,
args.n_emulator_runners, variables)
# Start all simulator processes
sim_coordinator.start()
network = QNetwork
def network_creator(name='value_learning', learning_network=None):
nonlocal args
args.name = name
return network(args, learning_network=learning_network)
learner = PDQFDLearner(network_creator, env_creator, args, sim_coordinator)
setup_kill_signal_handler(learner)
logger.info('Starting training')
learner.train()
logger.info('Finished training')
type=int,
help="Number of possible repetitions",
dest="nb_choices")
parser.add_argument('--checkpoint_interval',
default=1000000,
type=int,
help="Interval of steps btw checkpoints",
dest="checkpoint_interval")
parser.add_argument('--activation',
default='relu',
type=str,
help="activation function for the network",
dest="activation")
parser.add_argument('--alpha_leaky_relu',
default=0.1,
type=float,
help="coef for leaky relu",
dest="alpha_leaky_relu")
return parser
if __name__ == '__main__':
args = get_arg_parser().parse_args()
import logger_utils
logger_utils.save_args(args, args.debugging_folder)
logging.debug(args)
main(args)