max_path_length = 200
num_grad_updates = 1
use_maml = True
for v in variants:
task_var = v['task_var']
oracle = v['oracle']
if task_var == 0:
task_var = 'direc'
exp_prefix = 'bugfix_trpo_maml_antdirec' + str(max_path_length)
if oracle:
env = TfEnv(normalize(AntEnvDirecOracle()))
else:
env = TfEnv(normalize(AntEnvRandDirec()))
elif task_var == 1:
task_var = 'vel'
exp_prefix = 'posticml_trpo_maml_ant' + str(max_path_length)
if oracle:
env = TfEnv(normalize(AntEnvOracle()))
else:
env = TfEnv(normalize(AntEnvRand()))
elif task_var == 2:
print("HERE")
task_var = 'pos'
exp_prefix = 'posticml_trpo_maml_antpos_' + str(max_path_length)
if oracle:
env = TfEnv(normalize(AntEnvRandGoalOracle()))
else:
env = TfEnv(normalize(AntEnvRandGoal()))
step_sizes = [0.1, 0.2, 1.0, 0.0]
initial_params_files = [file1, file2, None, file3]
exp_names = [gen_name + name for name in names]
all_avg_returns = []
for step_i, initial_params_file in zip(range(len(step_sizes)), initial_params_files):
avg_returns = []
for goal_i, goal in zip(range(len(goals)), goals):
if initial_params_file is not None and 'oracle' in initial_params_file:
env = normalize(AntEnvDirecOracle())
n_itr = 1
else:
env = normalize(AntEnvRandDirec())
n_itr = 4
env = TfEnv(env)
policy = GaussianMLPPolicy( # random policy
name='policy',
env_spec=env.spec,
hidden_nonlinearity=tf.nn.relu,
hidden_sizes=(100, 100),
)
if initial_params_file is not None:
policy = None
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(
env=env,