def main():
env = TfEnv(CustomGymEnv('PointMazeLeft-v0'))
experts = load_latest_experts('data/point', n=50)
irl_model = GCLDiscrimTrajectory(env_spec=env.spec, expert_trajs=experts)
policy = GaussianMLPPolicy(name='policy', env_spec=env.spec, hidden_sizes=(32, 32))
algo = IRLTRPO(
env=env,
policy=policy,
irl_model=irl_model,
n_itr=2000,
batch_size=10000,
max_path_length=100,
discount=0.99,
store_paths=True,
discrim_train_itrs=50,
irl_model_wt=1.0,
entropy_weight=0.1, # this should be 1.0 but 0.1 seems to work better
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec)
)
with rllab_logdir(algo=algo, dirname='data/point_traj'):
with tf.Session():
algo.train()
test_pointmaze(sess.run(policy))
def main():
env = TfEnv(GymEnv('Ant-v1', record_video=False, record_log=False))
experts = load_latest_experts('data/ant', n=50)
irl_model = GCLDiscrim(
env_spec=env.spec,
expert_trajs=experts,
discrim_arch=disentangled_net)
policy = GaussianMLPPolicy(name='policy', env_spec=env.spec, hidden_sizes=(32, 32))
algo = IRLTRPO(
env=env,
policy=policy,
irl_model=irl_model,
n_itr=2000,
batch_size=10000,
max_path_length=1000,
discount=0.995,
store_paths=True,
discrim_train_itrs=50,
irl_model_wt=1.0,
entropy_weight=0.1,
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec)
)
with rllab_logdir(algo=algo, dirname='data/ant_airl'):
with tf.Session():
algo.train()
def main(num_examples=50, discount=0.99):
env = TfEnv(GymEnv('Pendulum-v0', record_video=False, record_log=False))
experts = load_latest_experts('data/pendulum', n=num_examples)
irl_model = GCLDiscrimTrajectory(env_spec=env.spec, expert_trajs=experts)
policy = GaussianMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=(32, 32))
algo = IRLTRPO(
env=env,
policy=policy,
irl_model=irl_model,
n_itr=200,
batch_size=2000,
max_path_length=100,
discount=discount,
store_paths=True,
discrim_train_itrs=50,
irl_model_wt=1.0,
entropy_weight=0.1, # this should be 1.0 but 0.1 seems to work better
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec))
with rllab_logdir(algo=algo, dirname='data/pendulum_traj'):
with tf.Session():
algo.train()
def main(exp_name=None, fusion=False, visible_gpus='0',discount=0.99):
env = TfEnv(CustomGymEnv('CustomAnt-v0', record_video=False, record_log=False))
gpu_options = tf.GPUOptions(allow_growth=True,visible_device_list=visible_gpus)
tf_config = tf.ConfigProto(inter_op_parallelism_threads=1, intra_op_parallelism_threads=1, gpu_options=gpu_options)
# load ~2 iterations worth of data from each forward RL experiment as demos
experts = load_latest_experts_multiple_runs('data/ant_data_collect', n=2, visible_gpus=visible_gpus)
irl_model = AIRL(discount=discount, env=env, expert_trajs=experts, state_only=True, fusion=fusion, max_itrs=10)
policy = GaussianMLPPolicy(name='policy', env_spec=env.spec, hidden_sizes=(32, 32))
algo = IRLTRPO(
env=env,
policy=policy,
irl_model=irl_model,
n_itr=1000,
batch_size=10000,
max_path_length=500,
discount=discount,
store_paths=True,
irl_model_wt=1.0,
entropy_weight=0.1,
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec),
)
with rllab_logdir(algo=algo, dirname='data/ant_state_irl/%s' % exp_name):
with tf.Session(config=tf_config) as sess:
algo.train(sess)
def main():
env = TfEnv(
GymEnv('HRI_AirSim_Landing-v0', record_video=False, record_log=False))
experts = load_latest_experts('data/airsim', n=5)
irl_model = AIRLStateAction(env_spec=env.spec, expert_trajs=experts)
policy = GaussianMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=(32, 32))
algo = IRLTRPO(
env=env,
policy=policy,
irl_model=irl_model,
n_itr=10,
batch_size=100,
max_path_length=100,
discount=0.99,
store_paths=True,
discrim_train_itrs=50,
irl_model_wt=1.0,
entropy_weight=0.1, # this should be 1.0 but 0.1 seems to work better
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec))
with rllab_logdir(algo=algo, dirname='data/airsim_gcl'):
with tf.Session():
algo.train()
def main(exp_name=None, fusion=False, visible_gpus='0', discount=0.99):
env = TfEnv(GymEnv('Swimmer-v3', record_video=False, record_log=False))
gpu_options = tf.GPUOptions(allow_growth=True,visible_device_list=args.visible_gpus)
tf_config = tf.ConfigProto(inter_op_parallelism_threads=1, intra_op_parallelism_threads=1, gpu_options=gpu_options)
experts = load_latest_experts('data/swimmer', n=5, visible_gpus=visible_gpus)
irl_model = AIRL(discount=discount, env=env, expert_trajs=experts, state_only=False, fusion=args.fusion, max_itrs=10)
policy = GaussianMLPPolicy(name='policy', env_spec=env.spec, hidden_sizes=(32, 32))
algo = IRLTRPO(
env=env,
policy=policy,
irl_model=irl_model,
n_itr=1000,
batch_size=10000,
max_path_length=1000,
discount=discount,
store_paths=True,
discrim_train_itrs=50,
irl_model_wt=1.0,
entropy_weight=0.1, # this should be 1.0 but 0.1 seems to work better
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec)
)
with rllab_logdir(algo=algo, dirname='data/swimmer_airl_state_action'):
with tf.Session(config=tf_config) as sess:
algo.train(sess)
def main(exp_name, params_folder=None):
env = TfEnv(CustomGymEnv('DisabledAnt-v0', record_video=False, record_log=False))
irl_itr = 100 # earlier IRL iterations overfit less; 100 seems to work well.
params_file = os.path.join(DATA_DIR, '%s/itr_%d.pkl' % (params_folder, irl_itr))
prior_params = load_prior_params(params_file)
irl_model = AIRL(env=env, expert_trajs=None, state_only=True)
policy = GaussianMLPPolicy(name='policy', env_spec=env.spec, hidden_sizes=(32, 32))
algo = IRLTRPO(
init_irl_params=prior_params,
env=env,
policy=policy,
irl_model=irl_model,
n_itr=1000,
batch_size=10000,
max_path_length=500,
discount=0.99,
store_paths=False,
train_irl=False,
irl_model_wt=1.0,
entropy_weight=0.1,
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec),
log_params_folder=params_folder,
log_experiment_name=exp_name,
)
with rllab_logdir(algo=algo, dirname='data/ant_transfer/%s'%exp_name):
with tf.Session():
algo.train()
def main():
env = TfEnv(
GymEnv('HRI_AirSim_Landing-v0', record_video=False, record_log=False))
### VGG 11/29/18: Added support to CSV files
## this method loads expert data saved as pickle file
# experts = load_latest_experts('data/airsim_final', n=1)
# this one uses csv:
experts = load_experts('data/airsim_human_data/log.csv',
pickle_format=False)
irl_model = GAIL(env_spec=env.spec, expert_trajs=experts)
policy = GaussianMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=(32, 32))
algo = IRLTRPO(
env=env,
policy=policy,
irl_model=irl_model,
n_itr=5000,
batch_size=60,
max_path_length=60,
discount=0.99,
store_paths=True,
discrim_train_itrs=100,
irl_model_wt=1.0,
entropy_weight=0.0, # GAIL should not use entropy unless for exploration
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec),
n_parallel=0)
with rllab_logdir(algo=algo, dirname='data/airsim_gail'):
with tf.Session():
algo.train()
def main():
env = TfEnv(GymEnv('Pendulum-v0', record_video=False, record_log=False))
experts = load_latest_experts('data/pendulum', n=5)
irl_model = GAIL(env_spec=env.spec, expert_trajs=experts)
policy = GaussianMLPPolicy(name='policy', env_spec=env.spec, hidden_sizes=(32, 32))
algo = IRLTRPO(
env=env,
policy=policy,
irl_model=irl_model,
n_itr=200,
batch_size=1000,
max_path_length=100,
discount=0.99,
store_paths=True,
discrim_train_itrs=50,
irl_model_wt=1.0,
entropy_weight=0.0, # GAIL should not use entropy unless for exploration
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec)
)
with rllab_logdir(algo=algo, dirname='data/pendulum_gail'):
with tf.Session():
algo.train()
def main(exp_name=None, fusion=False):
env = TfEnv(
CustomGymEnv('CustomAnt-v0', record_video=False, record_log=False))
# load ~2 iterations worth of data from each forward RL experiment as demos
experts = load_latest_experts_multiple_runs('data/ant_data_collect', n=2)
irl_model = AIRL(env=env,
expert_trajs=experts,
state_only=True,
fusion=fusion,
max_itrs=10)
policy = GaussianMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=(32, 32))
algo = IRLTRPO(
env=env,
policy=policy,
irl_model=irl_model,
n_itr=1000,
batch_size=10000,
max_path_length=500,
discount=0.99,
store_paths=True,
irl_model_wt=1.0,
entropy_weight=0.1,
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec),
)
with rllab_logdir(algo=algo, dirname='data/ant_state_irl/%s' % exp_name):
with tf.Session():
algo.train()
def main():
env = TfEnv(GymEnv('Pendulum-v0', record_video=False, record_log=False))
experts = load_latest_experts('data/pendulum', n=5)
irl_model = GAIL(env_spec=env.spec, expert_trajs=experts)
policy = GaussianMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=(32, 32))
algo = IRLTRPO(
env=env,
policy=policy,
irl_model=irl_model,
n_itr=200,
batch_size=1000,
max_path_length=100,
discount=0.99,
store_paths=True,
discrim_train_itrs=50,
irl_model_wt=1.0,
entropy_weight=0.0, # GAIL should not use entropy unless for exploration
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec))
with rllab_logdir(algo=algo, dirname='data/pendulum_gail'):
with tf.Session():
algo.train()
def main(eval_reward = False):
env = TfEnv(GymEnv('Pendulum-v0', record_video=False, record_log=False))
n_experts = 10
experts = load_latest_experts('plotting/pendulum_final', n=n_experts)
dirname='data/pendulum' # dir to save logs and images
irl_model = AIRLStateAction(env_spec=env.spec, expert_trajs=experts)
policy = GaussianMLPPolicy(name='policy', env_spec=env.spec, hidden_sizes=(32, 32))
algo = IRLTRPO(
env=env,
policy=policy,
irl_model=irl_model,
n_itr=1000,
batch_size=1000,
max_path_length=100,
discount=0.99,
store_paths=True,
discrim_train_itrs=50,
irl_model_wt=1.0,
entropy_weight=0.1, # this should be 1.0 but 0.1 seems to work better
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec),
eval_reward=True,
fig_dir = dirname
)
# with rllab_logdir(algo=algo, dirname='data/pendulum_gcl{}'.format(n_experts)):
with rllab_logdir(algo=algo, dirname=dirname):
with tf.Session():
algo.fig_dirname = dirname
algo.train()
def main(exp_name=None, fusion=True):
# env = TfEnv(CustomGymEnv('CustomAnt-v0', record_video=False, record_log=False))
env = TfEnv(
CustomGymEnv('CustomAnt-v0', record_video=False, record_log=True))
# load ~2 iterations worth of data from each forward RL experiment as demos
experts = load_latest_experts_multiple_runs('data/ant_data_collect', n=2)
#experts = load_latest_experts('data/ant_data_collect', n=5)
#qvar: inverse model q(a|s,s')
qvar = GaussianMLPInversePolicy(name='qvar_model',
env_spec=env.spec,
hidden_sizes=(32, 32))
qvar_model = Qvar(env=env,
qvar=qvar,
expert_trajs=experts,
fusion=True,
max_itrs=10)
#Empowerment-based Adversarial Inverse Reinforcement Learning, set score_discrim=True
irl_model = EAIRL(env=env,
expert_trajs=experts,
state_only=False,
fusion=fusion,
max_itrs=10,
score_discrim=True)
#Empowerment-based potential functions gamma* Phi(s')-Phi(s)
empw_model = Empowerment(env=env, fusion=True, max_itrs=4)
t_empw_model = Empowerment(env=env,
scope='t_efn',
fusion=True,
max_itrs=2,
name='empowerment2')
policy = GaussianMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=(32, 32))
algo = IRLTRPO(
env=env,
policy=policy,
empw=empw_model,
tempw=t_empw_model,
qvar_model=qvar_model,
irl_model=irl_model,
n_itr=3000, #130,
batch_size=20000,
max_path_length=500,
discount=0.99,
store_paths=True,
target_empw_update=5,
irl_model_wt=1.0,
entropy_weight=0.1,
lambda_i=1.0,
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec),
plot=False)
with rllab_logdir(algo=algo, dirname='data/ant_state_irl'):
#with rllab_logdir(algo=algo, dirname='data/ant_state_irl/%s' % exp_name): # if you use multiple runs, use this line instead of above
with tf.Session():
algo.train()
def main(exp_name=None, fusion=False, visible_gpus='0', discount=0.99, debug=False, n_val=1, n_rew=1, \
max_nstep=1, exp_folder=None, state_only=False, score_discrim=True, score_method=None):
env = TfEnv(
CustomGymEnv('PointMazeRight-v0', record_video=False,
record_log=False))
gpu_options = tf.GPUOptions(allow_growth=True,
visible_device_list=visible_gpus)
tf_config = tf.ConfigProto(inter_op_parallelism_threads=1,
intra_op_parallelism_threads=1,
gpu_options=gpu_options)
# load ~2 iterations worth of data from each forward RL experiment as demos
experts = load_latest_experts_multiple_runs('data/maze_right_data_collect',
n=2,
visible_gpus=visible_gpus)
sess = tf.Session(config=tf_config)
# sess = tf_debug.LocalCLIDebugWrapperSession(sess)
max_path_length = 500
irl_model = AIRL_Bootstrap(discount=discount, env=env, expert_trajs=experts, state_only=state_only, fusion=fusion, max_itrs=10, score_discrim=score_discrim, debug = debug, \
max_nstep = max_nstep, n_value_funct = n_val, n_rew_funct = n_rew, score_method=score_method)
policy = GaussianMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=(32, 32))
algo = IRLTRPO(
env=env,
policy=policy,
irl_model=irl_model,
n_itr=1000,
batch_size=10000,
max_path_length=max_path_length,
discount=discount,
store_paths=True,
irl_model_wt=1.0,
entropy_weight=0.1,
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec),
)
# temp_folder = '/media/data/temp_exp_nstep/maze_right_state_bootstrap_%d_irl/%s'
dirname = 'data/maze_right_state_bootstrap_%d_irl/%s/%s' % (
max_nstep, exp_folder, exp_name
) if exp_folder is not None else 'data/maze_right_state_bootstrap_%d_irl/%s' % (
max_nstep, exp_name)
with rllab_logdir(algo=algo, dirname=dirname):
sess.__enter__()
algo.train(sess)
sess.close()
def main(exp_name, params_folder=None, visible_gpus='0', discount=0.99):
env = TfEnv(
CustomGymEnv('PointMazeLeft-v0', record_video=False, record_log=False))
gpu_options = tf.GPUOptions(allow_growth=True,
visible_device_list=visible_gpus)
tf_config = tf.ConfigProto(inter_op_parallelism_threads=1,
intra_op_parallelism_threads=1,
gpu_options=gpu_options)
irl_itr = 100 # earlier IRL iterations overfit less; 100 seems to work well.
params_file = os.path.join(DATA_DIR,
'%s/itr_%d.pkl' % (params_folder, irl_itr))
prior_params = load_prior_params(params_file, tf_config)
irl_model = AIRL(discount=discount,
env=env,
expert_trajs=None,
state_only=True)
policy = GaussianMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=(32, 32))
algo = IRLTRPO(
init_irl_params=prior_params,
env=env,
policy=policy,
irl_model=irl_model,
n_itr=1000,
batch_size=10000,
max_path_length=500,
discount=discount,
store_paths=False,
train_irl=False,
irl_model_wt=1.0,
entropy_weight=0.1,
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec),
log_params_folder=params_folder,
log_experiment_name=exp_name,
)
with rllab_logdir(algo=algo,
dirname='data/maze_left_transfer/%s' % exp_name):
with tf.Session(config=tf_config) as sess:
algo.train(sess)
def main(exp_name=None, params_folder='data/ant_state_irl'):
# env = TfEnv(CustomGymEnv('PointMazeLeft-v0', record_video=True, record_log=True,force_reset=True))
env = TfEnv(
CustomGymEnv('DisabledAnt-v0',
record_video=False,
record_log=False,
force_reset=False))
irl_itr = 90 # earlier IRL iterations overfit less; either 80 or 90 seems to work well. But I usually search through 60,65,70,75, .. uptil 100
#params_file = os.path.join(DATA_DIR, '%s/itr_%d.pkl' % (params_folder, irl_itr))
params_file = os.path.join(DATA_DIR, 'itr_%d.pkl' % (irl_itr))
prior_params = load_prior_params(params_file)
'''q_itr = 400 # earlier IRL iterations overfit less; 100 seems to work well.
#params_file = os.path.join(DATA_DIR, '%s/itr_%d.pkl' % (params_folder, irl_itr))
params_file = os.path.join(DATA_DIR, 'itr_%d.pkl' % (q_itr))
prior_params_q = load_prior_params(params_file)'''
experts = load_latest_experts_multiple_runs('data/ant_data_collect', n=2)
qvar = GaussianMLPInversePolicy(name='qvar_model',
env_spec=env.spec,
hidden_sizes=(32, 32))
qvar_model = Qvar(env=env, qvar=qvar, expert_trajs=None, max_itrs=10)
irl_model = EAIRL(env=env,
expert_trajs=experts,
state_only=False,
score_discrim=False)
empw_model = Empowerment(env=env, max_itrs=1)
t_empw_model = Empowerment(env=env,
scope='t_efn',
max_itrs=2,
name='empowerment2')
policy = GaussianMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=(32, 32))
algo = IRLTRPO(
init_irl_params=prior_params['irl_params'],
init_empw_params=None, #prior_params['empw_params'],
init_qvar_params=None, #prior_params['qvar_params'],
init_policy_params=prior_params['policy_params'], #None
env=env,
policy=policy,
empw=empw_model,
tempw=t_empw_model,
qvar_model=qvar_model,
irl_model=irl_model,
n_itr=2000,
batch_size=20000,
max_path_length=500,
discount=0.99,
store_paths=False,
train_irl=True,
train_empw=True,
train_qvar=True,
irl_model_wt=1.0,
entropy_weight=0.1,
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec),
log_params_folder=params_folder,
log_experiment_name=exp_name,
# plot=True,
)
with rllab_logdir(algo=algo, dirname='data/ant_transfer'): #%s'%exp_name):
#with rllab_logdir(algo=algo, dirname='data/ant_transfer%s'%exp_name):
with tf.Session():
algo.train()
def main(exp_name,
rundir,
ent_wt=1.0,
env_name='Shaped_PM_MazeRoom_Small-v0',
method="airl",
beta=1e-2,
disc_step=1e-3,
disc_iters=100,
disc_batch_size=32,
disc_gp=None,
trpo_step=1e-2,
init_pol_std=1.0,
adaptive_beta=False,
target_kl=0.5,
beta_step=1e-6,
hid_size=None,
hid_layers=None,
max_traj=None):
os.makedirs(rundir, exist_ok=True)
if hid_size is None or hid_layers is None:
assert hid_size is None and hid_layers is None, \
"must specify both size & layers, not one or the other"
hid_layers, hid_size, init_pol_std \
= min_layers_hidsize_polstd_for(env_name)
env_trpo_params = irltrpo_params_for(env_name, 'irl')
env = TfEnv(CustomGymEnv(env_name, record_video=False, record_log=False))
expert_dir = os.path.join(rundir, 'env_%s/' % env_name.lower())
experts = load_latest_experts_walky(expert_dir, n=5, max_traj=max_traj)
disc_net_kwargs = {
'layers': hid_layers,
'd_hidden': hid_size,
}
if method in {'airl', 'vairl'}:
is_vairl = method == 'vairl'
irl_model = shaped_airl.AIRL(
env=env,
expert_trajs=experts,
state_only=True,
fusion=True,
discrim_arch_args=disc_net_kwargs,
fitted_value_fn_arch_args=disc_net_kwargs,
gp_coeff=disc_gp,
# vairl flag
vairl=is_vairl,
# vairl fixed beta settings
vairl_beta=beta,
# vairl adaptive beta settings
vairl_adaptive_beta=adaptive_beta,
vairl_beta_step_size=beta_step,
vairl_kl_target=target_kl)
elif method in {'gail', 'vail'}:
is_vail = method == 'vail'
assert gp_coeff is None, "no GAIL/VAIL support for GP coeff"
irl_model = GAIL(
env,
expert_trajs=experts,
discrim_arch_args=disc_net_kwargs,
name=method,
# vail stuff (only adaptive beta for VAIL, no fixed beta like
# VAIRL)
vail=is_vail,
# initial beta
vail_init_beta=beta,
vail_beta_step_size=beta_step,
vail_kl_target=target_kl)
else:
raise NotImplementedError("don't know how to handle method '%s'" %
(method, ))
pol_hid_sizes = (hid_size, ) * hid_layers
policy = GaussianMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=pol_hid_sizes,
init_std=init_pol_std)
irltrpo_kwargs = dict(
env=env,
policy=policy,
irl_model=irl_model,
discount=0.99,
store_paths=True,
discrim_train_itrs=disc_iters,
irl_model_wt=1.0,
irl_lr=disc_step,
irl_batch_size=disc_batch_size,
step_size=trpo_step,
# entropy_weight should be 1.0 but 0.1 seems to work better
entropy_weight=ent_wt,
force_batch_sampler=True,
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec),
)
irltrpo_kwargs.update(env_trpo_params)
n_itr = irltrpo_kwargs['n_itr']
print('irltrpo_kwargs:', irltrpo_kwargs)
algo = IRLTRPO(**irltrpo_kwargs)
run_name = 'env_{env_name}_{method}'.format(env_name=env_name.lower(),
method=method)
exp_folder = os.path.join(rundir, '%s/%s' % (run_name, exp_name))
with rllab_logdir(algo=algo, dirname=exp_folder):
with tf.Session():
algo.train()
this_dir = os.path.dirname(__file__)
maze_retrain_path = os.path.join(this_dir, 'env_retrain.py')
latest_irl_snap = '%s/itr_%d.pkl' % (exp_folder, n_itr - 1)
subproc_cmd = [
# script
'python',
maze_retrain_path,
# experiment info
latest_irl_snap,
'--rundir',
rundir,
# TRPO params
'--trpo-step',
'%f' % trpo_step,
'--trpo-ent',
'%f' % ent_wt,
# network params
'--hid-layers',
'%d' % hid_layers,
'--hid-size',
'%d' % hid_size,
# we don't care about precise args relevant to given method because
# we're just reloading a frozen model
'--method',
method,
]
subprocess.run(subproc_cmd, check=True)
trajectories.append(df)
experts.append(data_dict)
log.info("trajs : {}".format(len(trajectories)))
tf.reset_default_graph()
# if not osp.exists(gail_iter_path):
# os.makedirs(gail_iter_path)
# rllog.set_snapshot_dir(gail_iter_path)
with tf.Session():
env = TfEnv(env)
irl_model = GAIL(env_spec=env.spec, expert_trajs=experts)
policy = GaussianMLPPolicy(name='policy', env_spec=env.spec, hidden_sizes=(64, 64))
# policy._mean_network = iter_data['policy']._mean_network
algo = IRLTRPO(
env=env,
policy=policy,
irl_model=irl_model,
n_itr=t_iter,
batch_size=batch_size,
max_path_length=max_path_length,
discount=0.99,
store_paths=True,
discrim_train_itrs=75,
irl_model_wt=1.0,
entropy_weight=0.0, # GAIL should not use entropy unless for exploration
zero_environment_reward=True,
baseline=GaussianMLPBaseline(env_spec=env.spec)
)
algo.train()
# rllog.set_snapshot_dir(None)
def main(
exp_name,
rundir='data',
irl_pkl='',
ent_wt=1.0,
trpo_anneal_steps=None,
trpo_anneal_init_ent=None,
trpo_step=0.01,
init_pol_std=1.0,
method=None,
hid_size=None,
hid_layers=None,
switch_env=None,
):
orig_env_name = get_name(irl_pkl)
if switch_env is not None:
this_env_name = switch_env
else:
this_env_name = orig_env_name
print("Running on environment '%s'" % this_env_name)
env = TfEnv(
CustomGymEnv(this_env_name, record_video=False, record_log=False))
if hid_size is None or hid_layers is None:
assert hid_size is None and hid_layers is None, \
"must specify both size & layers, not one or the other"
hid_layers, hid_size, init_pol_std \
= min_layers_hidsize_polstd_for(orig_env_name)
env_trpo_params = irltrpo_params_for(orig_env_name, 'retrain')
folder = os.path.dirname(irl_pkl)
prior_params = load_prior_params(irl_pkl)
expert_dir = os.path.join(rundir, 'env_%s/' % orig_env_name.lower())
experts = load_latest_experts_walky(expert_dir, n=5)
# For some reason IRLTRPO is responsible for setting weights in this code.
# It would equally be possible to run global_variables_initializer()
# ourselves and then do irl_model.set_params(prior_params) if we just
# wanted to query energy, reward, etc. from the trained AIRL model without
# using IRLTRPO.
disc_net_kwargs = {
'layers': hid_layers,
'd_hidden': hid_size,
}
if method in {'airl', 'vairl'}:
irl_model = AIRL(env=env,
expert_trajs=experts,
state_only=True,
freeze=True,
vairl=method == 'vairl',
vairl_beta=1e-4,
discrim_arch_args=disc_net_kwargs,
fitted_value_fn_arch_args=disc_net_kwargs)
elif method in {'gail', 'vail'}:
irl_model = GAIL(env,
expert_trajs=experts,
discrim_arch_args=disc_net_kwargs,
name=method,
freeze=True,
vail=method == 'vail')
else:
raise NotImplementedError("Don't know how to handle method '%s'" %
method)
pol_hid_sizes = (hid_size, ) * hid_layers
policy = GaussianMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=pol_hid_sizes,
init_std=init_pol_std)
irltrpo_kwargs = dict(
env=env,
policy=policy,
irl_model=irl_model,
discount=0.99,
store_paths=True,
discrim_train_itrs=50,
irl_model_wt=1.0,
entropy_weight=ent_wt, # should be 1.0 but 0.1 seems to work better
step_size=trpo_step,
zero_environment_reward=True,
baseline=LinearFeatureBaseline(env_spec=env.spec),
init_irl_params=prior_params,
force_batch_sampler=True,
entropy_anneal_init_weight=trpo_anneal_init_ent,
entropy_anneal_steps=trpo_anneal_steps,
retraining=True)
irltrpo_kwargs.update(env_trpo_params)
algo = IRLTRPO(**irltrpo_kwargs)
folder_suffix = ''
if switch_env is not None:
# append lower case environment name to retrain folder path
folder_suffix = '_%s' % switch_env.lower()
with rllab_logdir(algo=algo,
dirname='%s/retrain%s' % (folder, folder_suffix)):
with tf.Session():
algo.train()
