def experiment(variant):
with open('expert_demos_listing.yaml', 'r') as f:
listings = yaml.load(f.read())
demos_path = listings[variant['expert_name']]['file_paths'][
variant['expert_idx']]
print(demos_path)
buffer_save_dict = joblib.load(demos_path)
target_state_buffer = buffer_save_dict['data']
# target_state_buffer /= variant['rescale']
state_indices = torch.LongTensor(variant['state_indices'])
env_specs = variant['env_specs']
env = get_env(env_specs)
env.seed(env_specs['eval_env_seed'])
print('\n\nEnv: {}'.format(env_specs['env_name']))
print('kwargs: {}'.format(env_specs['env_kwargs']))
print('Obs Space: {}'.format(env.observation_space))
print('Act Space: {}\n\n'.format(env.action_space))
policy = joblib.load(variant['policy_checkpoint'])['exploration_policy']
if variant['eval_deterministic']:
policy = MakeDeterministic(policy)
policy.to(ptu.device)
eval_sampler = PathSampler(env,
policy,
variant['num_eval_steps'],
variant['max_path_length'],
no_terminal=variant['no_terminal'],
render=variant['render'],
render_kwargs=variant['render_kwargs'])
test_paths = eval_sampler.obtain_samples()
obs = []
for path in test_paths:
obs += path['observations']
x = [o[0] for o in obs]
y = [o[1] for o in obs]
fig, ax = plt.subplots(figsize=(6, 6))
plt.scatter(x, y)
plt.xlim(-1.25, 20)
plt.ylim(-1.25, 10)
ax.set_yticks([0, 5, 10])
ax.set_xticks([0, 5, 10, 15, 20])
plt.savefig('./figs/' + variant['env_specs']['task_name'] + '.pdf',
bbox_inches='tight')
return 1
def experiment(variant):
env_specs = variant['env_specs']
env = get_env(env_specs)
env.seed(env_specs['eval_env_seed'])
print('\n\nEnv: {}'.format(env_specs['env_name']))
print('kwargs: {}'.format(env_specs['env_kwargs']))
print('Obs Space: {}'.format(env.observation_space))
print('Act Space: {}\n\n'.format(env.action_space))
if variant['scale_env_with_demo_stats']:
with open('expert_demos_listing.yaml', 'r') as f:
listings = yaml.load(f.read())
expert_demos_path = listings[variant['expert_name']]['file_paths'][
variant['expert_idx']]
buffer_save_dict = joblib.load(expert_demos_path)
env = ScaledEnv(
env,
obs_mean=buffer_save_dict['obs_mean'],
obs_std=buffer_save_dict['obs_std'],
acts_mean=buffer_save_dict['acts_mean'],
acts_std=buffer_save_dict['acts_std'],
)
policy = joblib.load(variant['policy_checkpoint'])['exploration_policy']
if variant['eval_deterministic']:
policy = MakeDeterministic(policy)
policy.to(ptu.device)
eval_sampler = PathSampler(env,
policy,
variant['num_eval_steps'],
variant['max_path_length'],
no_terminal=variant['no_terminal'],
render=variant['render'],
render_kwargs=variant['render_kwargs'])
test_paths = eval_sampler.obtain_samples()
average_returns = eval_util.get_average_returns(test_paths)
print(average_returns)
return 1
env.seed(env_specs['eval_env_seed'])
with open('expert_demos_listing.yaml', 'r') as f:
listings = yaml.load(f.read())
expert_demos_path = listings['norm_halfcheetah_32_demos_sub_20']['file_paths'][0]
buffer_save_dict = joblib.load(expert_demos_path)
env = ScaledEnv(
env,
obs_mean=buffer_save_dict['obs_mean'],
obs_std=buffer_save_dict['obs_std'],
acts_mean=buffer_save_dict['acts_mean'],
acts_std=buffer_save_dict['acts_std'],
)
bc_policy = joblib.load('/scratch/hdd001/home/kamyar/output/paper-version-hc-bc/paper_version_hc_bc_2019_05_19_00_32_05_0000--s-0/params.pkl')['exploration_policy']
bc_policy = MakeDeterministic(bc_policy)
bc_policy.to(ptu.device)
dagger_policy = joblib.load('/scratch/hdd001/home/kamyar/output/dagger-halfcheetah/dagger_halfcheetah_2019_08_20_16_30_36_0000--s-0/params.pkl')['exploration_policy']
dagger_policy = MakeDeterministic(dagger_policy)
dagger_policy.to(ptu.device)
irl_policy = joblib.load('/scratch/hdd001/home/kamyar/output/hc_airl_ckpt/params.pkl')['exploration_policy']
irl_policy = MakeDeterministic(irl_policy)
irl_policy.to(ptu.device)
fig, ax = plt.subplots(1)
eval_sampler = PathSampler(
env,
bc_policy,
20000,
def experiment(specs):
if not specs['use_scripted_policy']:
policy_is_scripted = False
policy = joblib.load(specs['expert_path'])['policy']
else:
policy_is_scripted = True
policy = get_scripted_policy(specs['scripted_policy_name'])
if specs['use_deterministic_expert']:
policy = MakeDeterministic(policy)
if ptu.gpu_enabled():
policy.to(ptu.device)
env = get_env(specs['env_specs'])
env.seed(specs['env_specs']['env_seed'])
# make the replay buffers
max_path_length = specs['max_path_length']
if 'wrap_absorbing' in specs and specs['wrap_absorbing']:
"""
There was an intial implementation for this in v1.0
in gen_irl_expert_trajs.py
"""
raise NotImplementedError()
_max_buffer_size = (max_path_length + 2) * specs['num_rollouts']
else:
_max_buffer_size = max_path_length * specs['num_rollouts']
_max_buffer_size = int(
np.ceil(_max_buffer_size / float(specs['subsample_factor'])))
buffer_constructor = lambda: EnvReplayBuffer(
_max_buffer_size,
env,
)
train_buffer = buffer_constructor()
test_buffer = buffer_constructor()
render = specs['render']
render_kwargs = specs['render_kwargs']
check_for_success = specs['check_for_success']
print('\n')
# fill the train buffer
fill_buffer(train_buffer,
env,
policy,
specs['num_rollouts'],
max_path_length,
no_terminal=specs['no_terminal'],
policy_is_scripted=policy_is_scripted,
render=render,
render_kwargs=render_kwargs,
check_for_success=check_for_success,
wrap_absorbing=False,
subsample_factor=specs['subsample_factor'])
# fill the test buffer
fill_buffer(test_buffer,
env,
policy,
specs['num_rollouts'],
max_path_length,
no_terminal=specs['no_terminal'],
policy_is_scripted=policy_is_scripted,
render=render,
render_kwargs=render_kwargs,
check_for_success=check_for_success,
wrap_absorbing=False,
subsample_factor=specs['subsample_factor'])
# save the replay buffers
logger.save_extra_data({
'train': train_buffer,
'test': test_buffer
},
name='expert_demos.pkl')
return 1