def train(num_timesteps, seed):
rank = MPI.COMM_WORLD.Get_rank()
sess = U.single_threaded_session()
sess.__enter__()
if rank == 0:
logger.configure()
else:
logger.configure(format_strs=[])
workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
set_global_seeds(workerseed)
config_file = os.path.join(os.path.dirname(os.path.realpath(__file__)), '..', 'configs',
'ant_navigate.yaml')
print(config_file)
env = AntNavigateEnv(config = config_file)
def policy_fn(name, ob_space, ac_space):
#return mlp_policy.MlpPolicy(name=name, ob_space=sensor_space, ac_space=ac_space, hid_size=64, num_hid_layers=2)
return cnn_policy.CnnPolicy(name=name, ob_space=ob_space, ac_space=ac_space, save_per_acts=10000, session=sess, kind='small')
env.seed(workerseed)
gym.logger.setLevel(logging.WARN)
pposgd_simple.learn(env, policy_fn,
max_timesteps=int(num_timesteps * 1.1 * 5),
timesteps_per_actorbatch=6000,
clip_param=0.2, entcoeff=0.00,
optim_epochs=4, optim_stepsize=1e-4, optim_batchsize=64,
gamma=0.99, lam=0.95,
schedule='linear',
save_per_acts=500
)
env.close()
def train(num_timesteps, seed):
rank = MPI.COMM_WORLD.Get_rank()
#sess = U.single_threaded_session()
sess = utils.make_gpu_session(args.num_gpu)
sess.__enter__()
if args.meta != "":
saver = tf.train.import_meta_graph(args.meta)
saver.restore(sess, tf.train.latest_checkpoint('./'))
if rank == 0:
logger.configure()
else:
logger.configure(format_strs=[])
workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
set_global_seeds(workerseed)
use_filler = not args.disable_filler
config_file = os.path.join(os.path.dirname(os.path.realpath(__file__)),
'..', 'configs', 'husky_navigate.yaml')
print(config_file)
raw_env = HuskyNavigateEnv(is_discrete=True,
gpu_count=args.gpu_count,
config=config_file)
# def policy_fn(name, ob_space, sensor_space, ac_space):
def policy_fn(name, ob_space, ac_space):
if args.mode == "SENSOR":
return mlp_policy.MlpPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space,
hid_size=64,
num_hid_layers=2)
else:
#return fuse_policy.FusePolicy(name=name, ob_space=ob_space, sensor_space=sensor_space, ac_space=ac_space, save_per_acts=10000, session=sess)
#else:
return cnn_policy.CnnPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space,
save_per_acts=10000,
session=sess,
kind='small')
env = Monitor(raw_env,
logger.get_dir() and osp.join(logger.get_dir(), str(rank)))
env.seed(workerseed)
gym.logger.setLevel(logging.WARN)
pposgd_simple.learn(env,
policy_fn,
max_timesteps=int(num_timesteps * 1.1),
timesteps_per_actorbatch=3000,
clip_param=0.2,
entcoeff=0.01,
optim_epochs=4,
optim_stepsize=3e-3,
optim_batchsize=64,
gamma=0.996,
lam=0.95,
schedule='linear',
save_name="husky_navigate_ppo_{}".format(args.mode),
save_per_acts=50,
sensor=args.mode == "SENSOR",
reload_name=args.reload_name)
'''
def train(seed):
rank = MPI.COMM_WORLD.Get_rank()
sess = utils.make_gpu_session(args.num_gpu)
sess.__enter__()
if args.meta != "":
saver = tf.train.import_meta_graph(args.meta)
saver.restore(sess, tf.train.latest_checkpoint('./'))
if rank == 0:
logger.configure()
else:
logger.configure(format_strs=[])
workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
set_global_seeds(workerseed)
use_filler = not args.disable_filler
config_file = os.path.join(os.path.dirname(os.path.realpath(__file__)),
'..', 'configs', 'config_husky.yaml')
print(config_file)
raw_env = HuskyNavigateEnv(gpu_idx=args.gpu_idx, config=config_file)
step = raw_env.config['n_step']
episode = raw_env.config['n_episode']
iteration = raw_env.config['n_iter']
elm_policy = raw_env.config['elm_active']
num_timesteps = step * episode * iteration
tpa = step * episode
if args.mode == "SENSOR": #Blind Mode
def policy_fn(name, ob_space, ac_space):
return mlp_policy.MlpPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space,
hid_size=128,
num_hid_layers=4,
elm_mode=elm_policy)
elif args.mode == "DEPTH" or args.mode == "RGB": #Fusing sensor space with image space
def policy_fn(name, ob_space, sensor_space, ac_space):
return fuse_policy.FusePolicy(name=name,
ob_space=ob_space,
sensor_space=sensor_space,
ac_space=ac_space,
save_per_acts=10000,
hid_size=128,
num_hid_layers=4,
session=sess,
elm_mode=elm_policy)
elif args.mode == "RESNET":
def policy_fn(name, ob_space, sensor_space, ac_space):
return resnet_policy.ResPolicy(name=name,
ob_space=ob_space,
sensor_space=sensor_space,
ac_space=ac_space,
save_per_acts=10000,
hid_size=128,
num_hid_layers=4,
session=sess,
elm_mode=elm_policy)
elif args.mode == "ODE":
def policy_fn(name, ob_space, sensor_space, ac_space):
return ode_policy.ODEPolicy(name=name,
ob_space=ob_space,
sensor_space=sensor_space,
ac_space=ac_space,
save_per_acts=10000,
hid_size=128,
num_hid_layers=4,
session=sess,
elm_mode=elm_policy)
else: #Using only image space
def policy_fn(name, ob_space, ac_space):
return cnn_policy.CnnPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space,
session=sess,
kind='small')
env = Monitor(raw_env,
logger.get_dir() and osp.join(logger.get_dir(), str(rank)))
env.seed(workerseed)
gym.logger.setLevel(logging.WARN)
args.reload_name = '/home/berk/PycharmProjects/Gibson_Exercise/gibson/utils/models/PPO_ODE_2020-12-05_500_50_137_150.model'
print(args.reload_name)
modes_camera = ["DEPTH", "RGB", "RESNET", "ODE"]
if args.mode in modes_camera:
pposgd_fuse.learn(env,
policy_fn,
max_timesteps=int(num_timesteps * 1.1),
timesteps_per_actorbatch=tpa,
clip_param=0.2,
entcoeff=0.03,
vfcoeff=0.01,
optim_epochs=4,
optim_stepsize=1e-3,
optim_batchsize=64,
gamma=0.99,
lam=0.95,
schedule='linear',
save_name="PPO_{}_{}_{}_{}_{}".format(
args.mode, datetime.date.today(), step, episode,
iteration),
save_per_acts=15,
reload_name=args.reload_name)
else:
if args.mode == "SENSOR": sensor = True
else: sensor = False
pposgd_simple.learn(env,
policy_fn,
max_timesteps=int(num_timesteps * 1.1),
timesteps_per_actorbatch=tpa,
clip_param=0.2,
entcoeff=0.03,
vfcoeff=0.01,
optim_epochs=4,
optim_stepsize=1e-3,
optim_batchsize=64,
gamma=0.996,
lam=0.95,
schedule='linear',
save_name="PPO_{}_{}_{}_{}_{}".format(
args.mode, datetime.date.today(), step,
episode, iteration),
save_per_acts=15,
sensor=sensor,
reload_name=args.reload_name)
env.close()
