def train(num_timesteps, seed):
rank = MPI.COMM_WORLD.Get_rank()
sess = utils.make_gpu_session(args.num_gpu)
sess.__enter__()
# 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)
config_file = os.path.join(os.path.dirname(os.path.realpath(__file__)), '..', 'configs',
'husky_gibson_flagrun_train.yaml')
print(config_file)
env = HuskyGibsonFlagRunEnv(config = config_file, gpu_count=args.gpu_count)
print(env.sensor_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=64, num_hid_layers=2, session=sess)
#env = bench.Monitor(env, logger.get_dir() and
# osp.join(logger.get_dir(), str(rank)))
#env.seed(workerseed)
gym.logger.setLevel(logging.WARN)
pposgd_fuse.learn(env, policy_fn,
max_timesteps=int(num_timesteps * 1.1),
timesteps_per_actorbatch=2048,
clip_param=0.2, entcoeff=0.01,
optim_epochs=4, optim_stepsize=1e-3, optim_batchsize=64,
gamma=0.99, lam=0.95,
schedule='linear',
save_name=args.save_name,
save_per_acts=50,
reload_name=args.reload_name
)
env.close()
def enjoy(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)
config_file = os.path.join(os.path.dirname(os.path.realpath(__file__)),
'..', 'configs',
'husky_gibson_flagrun_train_enjoy.yaml')
env = HuskyGibsonFlagRunEnv(gpu_idx=0, config=config_file)
step = env.config['n_step']
batch = env.config['n_batch']
iteration = env.config['n_iter']
num_timesteps = step * batch * iteration
tpa = step * batch
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,
hid_size=128,
num_hid_layers=2,
save_per_acts=10,
session=sess)
env.seed(workerseed)
gym.logger.setLevel(logging.WARN)
args.reload_name = '/home/berk/PycharmProjects/Gibson_Exercise/gibson/utils/models/PPO_DEPTH_2019-11-30_500_32_100_100.model'
pposgd_fuse.enjoy(
env,
policy_fn,
max_timesteps=int(num_timesteps * 1.1),
timesteps_per_actorbatch=tpa,
clip_param=0.2,
entcoeff=0.0,
optim_epochs=4,
optim_stepsize=3e-3,
optim_batchsize=64,
gamma=0.996,
lam=0.95,
schedule='linear',
save_per_acts=25,
reload_name=args.reload_name,
)
env.close()
def enjoy(num_timesteps, seed):
sess = utils.make_gpu_session(1)
sess.__enter__()
if args.meta != "":
saver = tf.train.import_meta_graph(args.meta)
saver.restore(sess, tf.train.latest_checkpoint('./'))
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',
'husky_gibson_flagrun_train.yaml')
env = HuskyGibsonFlagRunEnv(is_discrete=True,
gpu_count=0,
config=config_file)
#env = bench.Monitor(env, logger.get_dir() and
# osp.join(logger.get_dir(), str(rank)))
env.seed(workerseed)
# policy = cnn_policy.CnnPolicy(name='pi', ob_space=env.observation_space, ac_space=env.action_space, save_per_acts=10000, session=sess)
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,
session=sess)
policy = policy_fn("pi", env.observation_space, env.sensor_space,
env.action_space) # Construct network for new policy
reload_name = '/home/fei/Development/gibson/examples/train/models/flagrun_RGBD2_50.model'
if reload_name:
saver = tf.train.Saver()
saver.restore(tf.get_default_session(), reload_name)
print("Loaded model successfully.")
def execute_policy(env):
ob, ob_sensor = env.reset()
stochastic = False
while True:
# with Profiler("agent act"):
ac, vpred = policy.act(stochastic, ob, ob_sensor)
ob, rew, new, meta = env.step(ac)
ob_sensor = meta['sensor']
if new:
ob, ob_sensor = env.reset()
gym.logger.setLevel(logging.WARN)
#sess.run(init_op)
execute_policy(env)
env.close()
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('./'))
#sess = U.single_threaded_session()
#sess = utils.make_gpu_session(args.num_gpu)
#sess.__enter__()
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_gibson_flagrun_train.yaml')
print(config_file)
env = HuskyGibsonFlagRunEnv(config=config_file, gpu_idx=args.gpu_idx)
step = env.config['n_step']
batch = env.config['n_batch']
iteration = env.config['n_iter']
elm_policy = env.config['elm_active']
num_timesteps = step * batch * iteration
tpa = step * batch
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)
#env = bench.Monitor(env, logger.get_dir() and
# osp.join(logger.get_dir(), str(rank)))
args.reload_name = "/home/berk/PycharmProjects/Gibson_Exercise/gibson/utils/models/FLAG_DEPTH_2020-04-04_500_20_105_180.model"
print(args.reload_name)
env.seed(workerseed)
gym.logger.setLevel(logging.WARN)
pposgd_fuse.learn(env,
policy_fn,
max_timesteps=int(num_timesteps * 1.1),
timesteps_per_actorbatch=tpa,
clip_param=0.2,
entcoeff=0.01,
optim_epochs=4,
optim_stepsize=1e-3,
optim_batchsize=64,
gamma=0.99,
lam=0.95,
schedule='linear',
save_name="FLAG_{}_{}_{}_{}_{}".format(
args.mode, datetime.date.today(), step, batch,
iteration),
save_per_acts=15,
reload_name=args.reload_name)
env.close()