def main(args):
if args.dir is None:
print(
' ---------- Please mention current experiment directory ----------'
)
return
# Directory of current experiment
base_dir = os.path.dirname(os.path.realpath(__file__))
experiment_dir = os.path.join(base_dir, args.agent_type, args.dir)
# load traing/testing parameters
params = load_parameters(file=os.path.join(experiment_dir, 'params.dat'))
# print('env: ', params.environment)
# print('action_repeat: ', params.action_repeat)
# print('agent: ', params.agent)
# print('training_episodes: ', params.training_episodes)
# print('training_steps_per_episode: ', params.training_steps_per_episode)
# print('testing_episodes: ', params.testing_episodes)
# print('testing_steps_per_episode: ', params.testing_steps_per_episode)
# print('epsilon_start: ', params.hyperparameters['epsilon_start'])
# print('epsilon_end: ', params.hyperparameters['epsilon_end'])
# print('epsilon_decay: ', params.hyperparameters['epsilon_decay'])
# print('epsilon_steps: ', params.hyperparameters['epsilon_steps'])
# print('use_cuda: ', params.hyperparameters['use_cuda'])
# print('learning_rate: ', params.hyperparameters['learning_rate'])
# print('batch_size: ', params.hyperparameters['batch_size'])
# print('discount_rate: ', params.hyperparameters['discount_rate'])
# print('target_network_update_frequency: ', params.hyperparameters['target_network_update_frequency'])
# Initialize the environment
env = gym.make(params.environment)
state_size = env.observation_space.shape
action_size = env.action_space.n
# Initialize the agent
agent = DQNAgent(state_size=state_size,
action_size=action_size,
hyperparameters=params.hyperparameters)
if args.train:
trainer = Trainer(env=env,
agent=agent,
params=params,
exp_dir=experiment_dir)
try:
trainer.train()
except KeyboardInterrupt:
trainer.close()
sys.exit(0)
finally:
print('\ndone.')
if args.retrain:
trainer = Trainer(env=env,
agent=agent,
params=params,
exp_dir=experiment_dir,
retrain=True)
try:
trainer.retrain()
except KeyboardInterrupt:
trainer.close()
sys.exit(0)
finally:
print('\ndone.')
if args.test:
tester = Tester(env=env,
agent=agent,
params=params,
exp_dir=experiment_dir)
try:
tester.test()
except KeyboardInterrupt:
try:
tester.close()
sys.exit(0)
except SystemExit:
tester.close()
os._exit(0)
def main(args):
# Directory of current experiment
experiment_dir = 'experiments/dqn_lstm/test1'
# Load configuration
config = Config()
config.env = args.env
config.hyperparameters = {
"learning_rate": 0.025,
"batch_size": 32,
"sequence_length": 1,
"buffer_size": int(1e5),
"update_every_n_steps": 1,
"min_steps_before_learning": 1000,
"epsilon_start": 1,
"epsilon_end": 0.1,
"epsilon_decay": 0.995,
"discount_rate": 0.99,
"tau": 0.01,
}
config.use_cuda = True
config.number_of_episodes = 500
config.steps_per_episode = 500
config.previous_episode = 0
config.total_steps = 160000
config.pre_train_steps = 100
config.learing_frequency = 1
config.checkpoint = True
config.checkpoint_interval = 1
config.checkpoint_dir = experiment_dir + '/checkpoints'
config.log_dir = experiment_dir + '/logs'
config.model_dir = experiment_dir + '/model'
# Initialize the environment
env = gym.make('Urban-v0')
config.state_dim = env.observation_space.shape
config.action_dim = env.action_space.n
# Initialize the agent
agent = DDQNAgent(config)
# Initialize spawner
spawner = Spawner()
if args.train:
trainer = Trainer(env, agent, spawner, config)
try:
trainer.train()
except KeyboardInterrupt:
try:
trainer.close()
sys.exit(0)
except SystemExit:
trainer.close()
os._exit(0)
elif args.retrain:
if args.checkpoint_file is None:
print(
' ---------- Please mention checkoutpoint file name ----------'
)
return
trainer = Trainer(env, agent, spawner, config)
trainer.load_checkpoint(args.checkpoint_file)
try:
trainer.retrain()
except KeyboardInterrupt:
try:
trainer.close()
sys.exit(0)
except SystemExit:
trainer.close()
os._exit(0)
elif args.test:
tester = Tester(episodes, steps)
tester.load_checkpoint(args.checkpoint_file)
try:
tester.retrain()
except KeyboardInterrupt:
try:
tester.close()
sys.exit(0)
except SystemExit:
tester.close()
os._exit(0)