def setUpClass(cls):
env = GymEnv('Pendulum-v0')
random_pol = RandomPol(cls.env.observation_space, cls.env.action_space)
sampler = EpiSampler(cls.env, pol, num_parallel=1)
epis = sampler.sample(pol, max_steps=32)
traj = Traj()
traj.add_epis(epis)
traj.register_epis()
cls.num_step = traj.num_step
make_redis('localhost', '6379')
cls.r = get_redis()
cls.r.set('env', env)
cls.r.set('traj', traj)
pol_net = PolNet(env.observation_space, env.action_space)
gpol = GaussianPol(env.observation_space, env.action_space, pol_net)
pol_net = PolNet(env.observation_space,
env.action_space, deterministic=True)
dpol = DeterministicActionNoisePol(
env.observation_space, env.action_space, pol_net)
model_net = ModelNet(env.observation_space, env.action_space)
mpcpol = MPCPol(env.observation_space,
env.action_space, model_net, rew_func)
q_net = QNet(env.observation_space, env.action_space)
qfunc = DeterministicSAVfunc(
env.observation_space, env.action_space, q_net)
aqpol = ArgmaxQfPol(env.observation_space, env.action_space, qfunc)
v_net = VNet(env.observation_space)
vfunc = DeterministicSVfunc(env.observation_space, v_net)
cls.r.set('gpol', cloudpickle.dumps(gpol))
cls.r.set('dpol', cloudpickle.dumps(dpol))
cls.r.set('mpcpol', cloudpickle.dumps(mpcpol))
cls.r.set('qfunc', cloudpickle.dumps(qfunc))
cls.r.set('aqpol', cloudpickle.dumps(aqpol))
cls.r.set('vfunc', cloudpickle.dumps(vfunc))
c2d = C2DEnv(env)
pol_net = PolNet(c2d.observation_space, c2d.action_space)
mcpol = MultiCategoricalPol(
env.observation_space, env.action_space, pol_net)
cls.r.set('mcpol', cloudpickle.dumps(mcpol))
action_space = env.action_space
pol_net = PolNet(observation_space, action_space)
if isinstance(action_space, gym.spaces.Box):
pol = GaussianPol(observation_space,
action_space,
pol_net,
data_parallel=args.data_parallel)
elif isinstance(action_space, gym.spaces.Discrete):
pol = CategoricalPol(observation_space,
action_space,
pol_net,
data_parallel=args.data_parallel)
elif isinstance(action_space, gym.spaces.MultiDiscrete):
pol = MultiCategoricalPol(observation_space,
action_space,
pol_net,
data_parallel=args.data_parallel)
else:
raise ValueError('Only Box, Discrete, and MultiDiscrete are supported')
vf_net = VNet(observation_space)
vf = DeterministicSVfunc(observation_space,
vf_net,
data_parallel=args.data_parallel)
sampler = EpiSampler(env, pol, num_parallel=args.num_parallel, seed=args.seed)
optim_pol = torch.optim.Adam(pol_net.parameters(), args.pol_lr)
optim_vf = torch.optim.Adam(vf_net.parameters(), args.vf_lr)
with open(os.path.join(args.expert_dir, args.expert_fname), 'rb') as f:
h_size=256, cell_size=256)
s_pol_net = PolNetLSTM(observation_space, action_space,
h_size=256, cell_size=256)
else:
t_pol_net = PolNet(observation_space, action_space)
s_pol_net = PolNet(observation_space, action_space, h1=190, h2=90)
if isinstance(action_space, gym.spaces.Box):
t_pol = GaussianPol(observation_space, action_space, t_pol_net, args.rnn)
s_pol = GaussianPol(observation_space, action_space, s_pol_net, args.rnn)
elif isinstance(action_space, gym.spaces.Discrete):
t_pol = CategoricalPol(
observation_space, action_space, t_pol_net, args.rnn)
s_pol = CategoricalPol(
observation_space, action_space, s_pol_net, args.rnn)
elif isinstance(action_space, gym.spaces.MultiDiscrete):
t_pol = MultiCategoricalPol(
observation_space, action_space, t_pol_net, args.rnn)
s_pol = MultiCategoricalPol(
observation_space, action_space, s_pol_net, args.rnn)
else:
raise ValueError('Only Box, Discrete and Multidiscrete are supported')
if args.teacher_pol:
t_pol.load_state_dict(torch.load(
os.path.join(args.teacher_dir, args.teacher_fname)))
if args.rnn:
s_vf_net = VNetLSTM(observation_space, h_size=256, cell_size=256)
else:
s_vf_net = VNet(observation_space)
if args.sampling_policy == 'teacher':
# Please note that the two policies do not have to have the same hidden architecture
if args.rnn:
t_pol_net = PolNetLSTM(ob_space, ac_space, h_size=256, cell_size=256)
s_pol_net = PolNetLSTM(ob_space, ac_space, h_size=256, cell_size=256)
else:
t_pol_net = PolNet(ob_space, ac_space)
s_pol_net = PolNet(ob_space, ac_space, h1=190, h2=90)
if isinstance(ac_space, gym.spaces.Box):
t_pol = GaussianPol(ob_space, ac_space, t_pol_net, args.rnn)
s_pol = GaussianPol(ob_space, ac_space, s_pol_net, args.rnn)
elif isinstance(ac_space, gym.spaces.Discrete):
t_pol = CategoricalPol(ob_space, ac_space, t_pol_net, args.rnn)
s_pol = CategoricalPol(ob_space, ac_space, s_pol_net, args.rnn)
elif isinstance(ac_space, gym.spaces.MultiDiscrete):
t_pol = MultiCategoricalPol(ob_space, ac_space, t_pol_net, args.rnn)
s_pol = MultiCategoricalPol(ob_space, ac_space, s_pol_net, args.rnn)
else:
raise ValueError('Only Box, Discrete and Multidiscrete are supported')
if args.teacher_pol:
t_pol.load_state_dict(
torch.load(os.path.join(args.teacher_dir, args.teacher_fname)))
if args.rnn:
s_vf_net = VNetLSTM(ob_space, h_size=256, cell_size=256)
else:
s_vf_net = VNet(ob_space)
if args.sampling_policy == 'teacher':
teacher_sampler = EpiSampler(env,
ac_space,
pol_net,
args.rnn,
data_parallel=args.data_parallel,
parallel_dim=1 if args.rnn else 0)
elif isinstance(ac_space, gym.spaces.Discrete):
pol = CategoricalPol(ob_space,
ac_space,
pol_net,
args.rnn,
data_parallel=args.data_parallel,
parallel_dim=1 if args.rnn else 0)
elif isinstance(ac_space, gym.spaces.MultiDiscrete):
pol = MultiCategoricalPol(ob_space,
ac_space,
pol_net,
args.rnn,
data_parallel=args.data_parallel,
parallel_dim=1 if args.rnn else 0)
else:
raise ValueError('Only Box, Discrete, and MultiDiscrete are supported')
if args.rnn:
vf_net = VNetLSTM(ob_space, h_size=256, cell_size=256)
else:
vf_net = VNet(ob_space)
vf = DeterministicSVfunc(ob_space,
vf_net,
args.rnn,
data_parallel=args.data_parallel,
parallel_dim=1 if args.rnn else 0)
def main(args):
init_ray(args.num_cpus, args.num_gpus, args.ray_redis_address)
if not os.path.exists(args.log):
os.makedirs(args.log)
if not os.path.exists(os.path.join(args.log, 'models')):
os.mkdir(os.path.join(args.log, 'models'))
score_file = os.path.join(args.log, 'progress.csv')
logger.add_tabular_output(score_file)
logger.add_tensorboard_output(args.log)
with open(os.path.join(args.log, 'args.json'), 'w') as f:
json.dump(vars(args), f)
pprint(vars(args))
# when doing the distributed training, disable video recordings
env = GymEnv(args.env_name)
env.env.seed(args.seed)
if args.c2d:
env = C2DEnv(env)
observation_space = env.observation_space
action_space = env.action_space
pol_net = PolNet(observation_space, action_space)
rnn = False
# pol_net = PolNetLSTM(observation_space, action_space)
# rnn = True
if isinstance(action_space, gym.spaces.Box):
pol = GaussianPol(observation_space, action_space, pol_net, rnn=rnn)
elif isinstance(action_space, gym.spaces.Discrete):
pol = CategoricalPol(observation_space, action_space, pol_net)
elif isinstance(action_space, gym.spaces.MultiDiscrete):
pol = MultiCategoricalPol(observation_space, action_space, pol_net)
else:
raise ValueError('Only Box, Discrete, and MultiDiscrete are supported')
vf_net = VNet(observation_space)
vf = DeterministicSVfunc(observation_space, vf_net)
trainer = TrainManager(Trainer,
args.num_trainer,
args.master_address,
args=args,
vf=vf,
pol=pol)
sampler = EpiSampler(env, pol, args.num_parallel, seed=args.seed)
total_epi = 0
total_step = 0
max_rew = -1e6
start_time = time.time()
while args.max_epis > total_epi:
with measure('sample'):
sampler.set_pol_state(trainer.get_state("pol"))
epis = sampler.sample(max_steps=args.max_steps_per_iter)
with measure('train'):
result_dict = trainer.train(epis=epis)
step = result_dict["traj_num_step"]
total_step += step
total_epi += result_dict["traj_num_epi"]
rewards = [np.sum(epi['rews']) for epi in epis]
mean_rew = np.mean(rewards)
elapsed_time = time.time() - start_time
logger.record_tabular('ElapsedTime', elapsed_time)
logger.record_results(args.log,
result_dict,
score_file,
total_epi,
step,
total_step,
rewards,
plot_title=args.env_name)
with measure('save'):
pol_state = trainer.get_state("pol")
vf_state = trainer.get_state("vf")
optim_pol_state = trainer.get_state("optim_pol")
optim_vf_state = trainer.get_state("optim_vf")
torch.save(pol_state,
os.path.join(args.log, 'models', 'pol_last.pkl'))
torch.save(vf_state, os.path.join(args.log, 'models',
'vf_last.pkl'))
torch.save(optim_pol_state,
os.path.join(args.log, 'models', 'optim_pol_last.pkl'))
torch.save(optim_vf_state,
os.path.join(args.log, 'models', 'optim_vf_last.pkl'))
if mean_rew > max_rew:
torch.save(pol_state,
os.path.join(args.log, 'models', 'pol_max.pkl'))
torch.save(vf_state,
os.path.join(args.log, 'models', 'vf_max.pkl'))
torch.save(
optim_pol_state,
os.path.join(args.log, 'models', 'optim_pol_max.pkl'))
torch.save(
optim_vf_state,
os.path.join(args.log, 'models', 'optim_vf_max.pkl'))
max_rew = mean_rew
del sampler
del trainer
log_dir=os.path.join(args.log, 'movie'),
record_video=args.record)
env.env.seed(args.seed)
if args.c2d:
env = C2DEnv(env)
observation_space = env.observation_space
action_space = env.action_space
pol_net = PolNet(observation_space, action_space)
if isinstance(action_space, gym.spaces.Box):
pol = GaussianPol(observation_space, action_space, pol_net)
elif isinstance(action_space, gym.spaces.Discrete):
pol = CategoricalPol(observation_space, action_space, pol_net)
elif isinstance(action_space, gym.spaces.MultiDiscrete):
pol = MultiCategoricalPol(observation_space, action_space, pol_net)
else:
raise ValueError('Only Box, Discrete, and MultiDiscrete are supported')
sampler = EpiSampler(env, pol, num_parallel=args.num_parallel, seed=args.seed)
optim_pol = torch.optim.Adam(pol_net.parameters(), args.pol_lr)
with open(os.path.join(args.expert_dir, args.expert_fname), 'rb') as f:
expert_epis = pickle.load(f)
train_epis, test_epis = ef.train_test_split(expert_epis,
train_size=args.train_size)
train_traj = Traj()
train_traj.add_epis(train_epis)
train_traj.register_epis()
test_traj = Traj()
test_traj.add_epis(test_epis)
if args.c2d:
env = C2DEnv(env)
ob_space = env.observation_space
ac_space = env.action_space
pol_net = PolNet(ob_space, ac_space)
if isinstance(ac_space, gym.spaces.Box):
pol = GaussianPol(ob_space, ac_space, pol_net,
data_parallel=args.data_parallel)
elif isinstance(ac_space, gym.spaces.Discrete):
pol = CategoricalPol(ob_space, ac_space, pol_net,
data_parallel=args.data_parallel)
elif isinstance(ac_space, gym.spaces.MultiDiscrete):
pol = MultiCategoricalPol(
ob_space, ac_space, pol_net, data_parallel=args.data_parallel)
else:
raise ValueError('Only Box, Discrete, and MultiDiscrete are supported')
vf_net = VNet(ob_space)
vf = DeterministicSVfunc(ob_space, vf_net,
data_parallel=args.data_parallel)
if args.rew_type == 'rew':
rewf_net = VNet(ob_space, h1=args.discrim_h1, h2=args.discrim_h2)
rewf = DeterministicSVfunc(
ob_space, rewf_net, data_parallel=args.data_parallel)
shaping_vf_net = VNet(ob_space, h1=args.discrim_h1, h2=args.discrim_h2)
shaping_vf = DeterministicSVfunc(
ob_space, shaping_vf_net, data_parallel=args.data_parallel)
optim_discrim = torch.optim.Adam(
env2 = C2DEnv(env2)
assert env1.ob_space == env2.ob_space
assert env1.ac_space.shape == env2.ac_space.shape
ob_space = env1.observation_space
ac_space = env1.action_space
pol_net = PolNetLSTM(ob_space,
ac_space,
h_size=args.h_size,
cell_size=args.cell_size)
pol = MultiCategoricalPol(ob_space,
ac_space,
pol_net,
True,
data_parallel=args.data_parallel,
parallel_dim=1)
vf_net = VNetLSTM(ob_space, h_size=args.h_size, cell_size=args.cell_size)
vf = DeterministicSVfunc(ob_space,
vf_net,
True,
data_parallel=args.data_parallel,
parallel_dim=1)
sampler1 = EpiSampler(env1,
pol,
num_parallel=args.num_parallel,
seed=args.seed)
sampler2 = EpiSampler(env2,
action_space,
pol_net,
args.rnn,
data_parallel=args.ddp,
parallel_dim=1 if args.rnn else 0)
elif isinstance(action_space, gym.spaces.Discrete):
pol = CategoricalPol(observation_space,
action_space,
pol_net,
args.rnn,
data_parallel=args.ddp,
parallel_dim=1 if args.rnn else 0)
elif isinstance(action_space, gym.spaces.MultiDiscrete):
pol = MultiCategoricalPol(observation_space,
action_space,
pol_net,
args.rnn,
data_parallel=args.ddp,
parallel_dim=1 if args.rnn else 0)
else:
raise ValueError('Only Box, Discrete, and MultiDiscrete are supported')
if args.rnn:
vf_net = VNetLSTM(observation_space, h_size=256, cell_size=256)
else:
vf_net = VNet(observation_space)
vf = DeterministicSVfunc(observation_space,
vf_net,
args.rnn,
data_parallel=args.ddp,
parallel_dim=1 if args.rnn else 0)