hidden_nonlinearity=tf.nn.relu,
hidden_sizes=(100, 100),
)
if bas == 'zero':
baseline = ZeroBaseline(env_spec=env.spec)
elif bas == 'linear':
baseline = LinearFeatureBaseline(env_spec=env.spec)
else:
baseline = GaussianMLPBaseline(env_spec=env.spec)
algo = SensitiveTRPO(
#algo = SensitiveVPG(
env=env,
policy=policy,
baseline=baseline,
batch_size=fast_batch_size, # number of trajs for grad update
max_path_length=max_path_length,
meta_batch_size=meta_batch_size,
num_grad_updates=num_grad_updates,
n_itr=400,
use_sensitive=use_sensitive,
#optimizer_args={'tf_optimizer_args':{'learning_rate': learning_rate}},
plot=False,
)
run_experiment_lite(
algo.train(),
n_parallel=0,
snapshot_mode="last",
seed=1,
#exp_prefix='deleteme',
#exp_name='deleteme'
#exp_prefix='sensitive1dT5_2017_01_19',
#exp_prefix='bugfix_sensitive0d_8tasks_T'+str(max_path_length)+'_2017_02_05',
hidden_nonlinearity=tf.nn.relu,
hidden_sizes=(100, 100),
)
if bas == 'zero':
baseline = ZeroBaseline(env_spec=env.spec)
elif 'linear' in bas:
baseline = LinearFeatureBaseline(env_spec=env.spec)
else:
baseline = GaussianMLPBaseline(env_spec=env.spec)
algo = SensitiveTRPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=fast_batch_size, # number of trajs for grad update
max_path_length=max_path_length,
meta_batch_size=meta_batch_size,
num_grad_updates=num_grad_updates,
n_itr=100,
use_sensitive=use_sensitive,
step_size=meta_step_size,
plot=False,
)
run_experiment_lite(
algo.train(),
n_parallel=40,
snapshot_mode="last",
python_command='python3',
seed=1,
exp_prefix='vpg_sensitive_point100',
exp_name='trposens' + str(int(use_sensitive)) + '_fbs' +
str(fast_batch_size) + '_mbs' + str(meta_batch_size) +