def experiment(variant):
task_mode = variant['task_mode'] # train, test, eval
task_idx = variant['task_idx']
if task_mode == 'train':
task_sampler = WalkerTrainParamsSampler()
elif task_mode == 'test':
task_sampler = WalkerTestParamsSampler()
else:
raise NotImplementedError()
task_params = task_sampler.get_task(task_idx)
obs_task_params = task_sampler.get_obs_task_params(task_params)
env = SingleTaskWalkerEnv(task_params, obs_task_params)
training_env = SingleTaskWalkerEnv(task_params, obs_task_params)
print(env.observation_space)
obs_dim = int(np.prod(env.observation_space.shape))
action_dim = int(np.prod(env.action_space.shape))
net_size = variant['net_size']
hidden_sizes = [net_size] * variant['num_hidden_layers']
print('Using simple model')
qf1 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + action_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + action_dim,
output_size=1,
)
vf = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=hidden_sizes,
obs_dim=obs_dim,
action_dim=action_dim,
)
algorithm = NewSoftActorCritic(
env=env,
training_env=training_env,
policy=policy,
qf1=qf1,
qf2=qf2,
vf=vf,
**variant['algo_params']
)
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
return 1
def experiment(variant):
with open(EXPERT_LISTING_YAML_PATH, 'r') as f:
listings = yaml.load(f.read())
expert_dir = listings[variant['expert_name']]['exp_dir']
specific_run = listings[variant['expert_name']]['seed_runs'][
variant['expert_seed_run_idx']]
file_to_load = path.join(expert_dir, specific_run, 'extra_data.pkl')
extra_data = joblib.load(file_to_load)
# this script is for the non-meta-learning airl
train_context_buffer, train_test_buffer = extra_data['meta_train'][
'context'], extra_data['meta_train']['test']
test_context_buffer, test_test_buffer = extra_data['meta_test'][
'context'], extra_data['meta_test']['test']
# load the expert
expert_policy = joblib.load(variant['expert_policy'])['algorithm']
expert_policy.replay_buffer = None
# set up the envs
env_specs = variant['env_specs']
meta_train_env, meta_test_env = get_meta_env(env_specs)
meta_train_env.seed(variant['seed'])
meta_test_env.seed(variant['seed'])
if variant['scale_env_with_given_demo_stats']:
assert not env_specs['normalized']
meta_train_env = ScaledMetaEnv(
meta_train_env,
obs_mean=extra_data['obs_mean'],
obs_std=extra_data['obs_std'],
acts_mean=extra_data['acts_mean'],
acts_std=extra_data['acts_std'],
)
meta_test_env = ScaledMetaEnv(
meta_test_env,
obs_mean=extra_data['obs_mean'],
obs_std=extra_data['obs_std'],
acts_mean=extra_data['acts_mean'],
acts_std=extra_data['acts_std'],
)
print(meta_train_env)
print(meta_test_env)
# set up the policy and training algorithm
if isinstance(meta_train_env.observation_space, Dict):
if variant['algo_params']['policy_uses_pixels']:
raise NotImplementedError(
'Not implemented pixel version of things!')
else:
obs_dim = int(
np.prod(meta_train_env.observation_space.spaces['obs'].shape))
else:
obs_dim = int(np.prod(meta_train_env.observation_space.shape))
action_dim = int(np.prod(meta_train_env.action_space.shape))
print('obs dim: %d' % obs_dim)
print('act dim: %d' % action_dim)
sleep(3)
# make the disc model
z_dim = variant['algo_params']['z_dim']
policy_net_size = variant['policy_net_size']
hidden_sizes = [policy_net_size] * variant['num_hidden_layers']
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=hidden_sizes,
obs_dim=obs_dim + z_dim,
action_dim=action_dim,
)
# Make the encoder
encoder = TimestepBasedEncoder(
2 * obs_dim + action_dim, #(s,a,s')
variant['algo_params']['r_dim'],
variant['algo_params']['z_dim'],
variant['algo_params']['enc_hid_dim'],
variant['algo_params']['r2z_hid_dim'],
variant['algo_params']['num_enc_layer_blocks'],
hid_act='relu',
use_bn=True,
within_traj_agg=variant['algo_params']['within_traj_agg'])
train_task_params_sampler, test_task_params_sampler = get_meta_env_params_iters(
env_specs)
algorithm = MetaDagger(
meta_test_env, # env is the test env, training_env is the training env (following rlkit original setup)
policy,
expert_policy,
train_context_buffer,
train_test_buffer,
test_context_buffer,
test_test_buffer,
encoder,
training_env=meta_train_env, # the env used for generating trajectories
train_task_params_sampler=train_task_params_sampler,
test_task_params_sampler=test_task_params_sampler,
**variant['algo_params'])
for task_id in train_context_buffer.task_replay_buffers:
erb = train_context_buffer.task_replay_buffers[task_id]
rb = algorithm.replay_buffer.task_replay_buffers[task_id]
erb_size = erb._size
print(erb_size)
for k in erb._observations:
rb._observations[k][:erb_size] = erb._observations[k][:erb_size]
rb._next_obs[k][:erb_size] = erb._next_obs[k][:erb_size]
rb._actions[:erb_size] = erb._actions[:erb_size]
rb._rewards[:erb_size] = erb._rewards[:erb_size]
rb._terminals[:erb_size] = erb._terminals[:erb_size]
rb._absorbing[:erb_size] = erb._absorbing[:erb_size]
rb._size = erb_size
rb._top = erb_size
# print('\n\n')
# for task_id in algorithm.replay_buffer.task_replay_buffers:
# rb = algorithm.replay_buffer.task_replay_buffers[task_id]
# print(rb._size)
# print(rb._top)
# print(rb._max_replay_buffer_size)
if ptu.gpu_enabled():
expert_policy.cuda()
algorithm.cuda()
algorithm.train()
return 1
def experiment(variant):
# we have to generate the combinations for the env_specs
env_specs = variant['env_specs']
env_specs_vg = VariantGenerator()
env_spec_constants = {}
env_spec_ranges = {}
for k, v in env_specs.items():
if isinstance(v, list):
env_specs_vg.add(k, v)
env_spec_ranges[k] = v
else:
env_spec_constants[k] = v
env_specs_list = []
for es in env_specs_vg.variants():
del es['_hidden_keys']
es.update(env_spec_constants)
env_specs_list.append(es)
env_sampler = EnvSampler(env_specs_list)
# make the normalizer function for the env_params
mean = []
half_diff = []
for k in sorted(env_spec_ranges.keys()):
r = env_spec_ranges[k]
if len(r) == 1:
mean.append(0)
half_diff.append(r[0])
else:
mean.append((r[0] + r[1]) / 2.0)
half_diff.append((r[1] - r[0]) / 2.0)
mean = np.array(mean)
half_diff = np.array(half_diff)
def env_params_normalizer(params):
return (params - mean) / half_diff
variant['algo_params']['env_params_normalizer'] = env_params_normalizer
# set up similar to non-meta version
sample_env, _ = env_sampler()
if variant['algo_params']['concat_env_params_to_obs']:
meta_params_dim = sample_env.env_meta_params.shape[0]
else:
meta_params_dim = 0
obs_dim = int(np.prod(sample_env.observation_space.shape))
action_dim = int(np.prod(sample_env.action_space.shape))
net_size = variant['net_size']
vf = FlattenMlp(
hidden_sizes=[net_size, net_size],
input_size=obs_dim + meta_params_dim,
output_size=1,
)
if exp_specs['use_new_sac']:
qf1 = FlattenMlp(
hidden_sizes=[net_size, net_size],
input_size=obs_dim + action_dim + meta_params_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=[net_size, net_size],
input_size=obs_dim + action_dim + meta_params_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=[net_size, net_size],
obs_dim=obs_dim + meta_params_dim,
action_dim=action_dim,
)
algorithm = NewMetaSoftActorCritic(env_sampler=env_sampler,
policy=policy,
qf1=qf1,
qf2=qf2,
vf=vf,
**variant['algo_params'])
else:
policy = TanhGaussianPolicy(
hidden_sizes=[net_size, net_size],
obs_dim=obs_dim + meta_params_dim,
action_dim=action_dim,
)
qf = FlattenMlp(
hidden_sizes=[net_size, net_size],
input_size=obs_dim + action_dim + meta_params_dim,
output_size=1,
)
algorithm = MetaSoftActorCritic(env_sampler=env_sampler,
policy=policy,
qf=qf,
vf=vf,
**variant['algo_params'])
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
return 1
def experiment(variant):
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)
expert_replay_buffer = buffer_save_dict['train']
env_specs = variant['env_specs']
env = get_env(env_specs)
env.seed(env_specs['eval_env_seed'])
training_env = get_env(env_specs)
training_env.seed(env_specs['training_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']:
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'],
)
training_env = ScaledEnv(
training_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'],
)
obs_space = env.observation_space
act_space = env.action_space
assert not isinstance(obs_space, Dict)
assert len(obs_space.shape) == 1
assert len(act_space.shape) == 1
obs_dim = obs_space.shape[0]
action_dim = act_space.shape[0]
# build the policy models
net_size = variant['policy_net_size']
num_hidden = variant['policy_num_hidden_layers']
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=num_hidden * [net_size],
obs_dim=obs_dim,
action_dim=action_dim,
)
algorithm = BC(env=env,
training_env=training_env,
exploration_policy=policy,
expert_replay_buffer=expert_replay_buffer,
**variant['bc_params'])
if ptu.gpu_enabled():
algorithm.to(ptu.device)
algorithm.train()
return 1
def experiment(variant):
with open(EXPERT_LISTING_YAML_PATH, 'r') as f:
listings = yaml.load(f.read())
expert_dir = listings[variant['expert_name']]['exp_dir']
specific_run = listings[variant['expert_name']]['seed_runs'][variant['expert_seed_run_idx']]
file_to_load = path.join(expert_dir, specific_run, 'extra_data.pkl')
extra_data = joblib.load(file_to_load)
# this script is for the non-meta-learning airl
train_context_buffer, train_test_buffer = extra_data['meta_train']['context'], extra_data['meta_train']['test']
test_context_buffer, test_test_buffer = extra_data['meta_test']['context'], extra_data['meta_test']['test']
# set up the envs
env_specs = variant['env_specs']
meta_train_env, meta_test_env = get_meta_env(env_specs)
meta_train_env.seed(variant['seed'])
meta_test_env.seed(variant['seed'])
if variant['scale_env_with_given_demo_stats']:
assert not env_specs['normalized']
meta_train_env = ScaledMetaEnv(
meta_train_env,
obs_mean=extra_data['obs_mean'],
obs_std=extra_data['obs_std'],
acts_mean=extra_data['acts_mean'],
acts_std=extra_data['acts_std'],
)
meta_test_env = ScaledMetaEnv(
meta_test_env,
obs_mean=extra_data['obs_mean'],
obs_std=extra_data['obs_std'],
acts_mean=extra_data['acts_mean'],
acts_std=extra_data['acts_std'],
)
print(meta_train_env)
print(meta_test_env)
# set up the policy and training algorithm
if isinstance(meta_train_env.observation_space, Dict):
if variant['algo_params']['policy_uses_pixels']:
raise NotImplementedError('Not implemented pixel version of things!')
else:
obs_dim = int(np.prod(meta_train_env.observation_space.spaces['obs'].shape))
else:
obs_dim = int(np.prod(meta_train_env.observation_space.shape))
action_dim = int(np.prod(meta_train_env.action_space.shape))
print('obs dim: %d' % obs_dim)
print('act dim: %d' % action_dim)
sleep(3)
# make the disc model
z_dim = variant['algo_params']['z_dim']
policy_net_size = variant['policy_net_size']
hidden_sizes = [policy_net_size] * variant['num_hidden_layers']
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=hidden_sizes,
obs_dim=obs_dim + z_dim,
action_dim=action_dim,
)
# Make the encoder
encoder = TimestepBasedEncoder(
2*obs_dim + action_dim, #(s,a,s')
variant['algo_params']['r_dim'],
variant['algo_params']['z_dim'],
variant['algo_params']['enc_hid_dim'],
variant['algo_params']['r2z_hid_dim'],
variant['algo_params']['num_enc_layer_blocks'],
hid_act='relu',
use_bn=True,
within_traj_agg=variant['algo_params']['within_traj_agg']
)
# ---------------
# encoder = WeightShareTimestepBasedEncoder(
# obs_dim,
# action_dim,
# 64,
# variant['algo_params']['r_dim'],
# variant['algo_params']['z_dim'],
# variant['algo_params']['enc_hid_dim'],
# variant['algo_params']['r2z_hid_dim'],
# variant['algo_params']['num_enc_layer_blocks'],
# hid_act='relu',
# use_bn=True,
# within_traj_agg=variant['algo_params']['within_traj_agg']
# )
# ---------------
# traj_enc = ConvTrajEncoder(
# variant['algo_params']['np_params']['traj_enc_params']['num_conv_layers'],
# # obs_dim + action_dim,
# obs_dim + action_dim + obs_dim,
# variant['algo_params']['np_params']['traj_enc_params']['channels'],
# variant['algo_params']['np_params']['traj_enc_params']['kernel'],
# variant['algo_params']['np_params']['traj_enc_params']['stride'],
# )
# Dc2R_map = Dc2RMap(
# variant['algo_params']['np_params']['Dc2r_params']['agg_type'],
# traj_enc,
# state_only=False
# )
# r2z_map = R2ZMap(
# variant['algo_params']['np_params']['r2z_params']['num_layers'],
# variant['algo_params']['np_params']['traj_enc_params']['channels'],
# variant['algo_params']['np_params']['r2z_params']['hid_dim'],
# variant['algo_params']['z_dim']
# )
# encoder = NPEncoder(
# Dc2R_map,
# r2z_map,
# )
train_task_params_sampler, test_task_params_sampler = get_meta_env_params_iters(env_specs)
algorithm = NeuralProcessBC(
meta_test_env, # env is the test env, training_env is the training env (following rlkit original setup)
policy,
train_context_buffer,
train_test_buffer,
test_context_buffer,
test_test_buffer,
encoder,
training_env=meta_train_env, # the env used for generating trajectories
train_task_params_sampler=train_task_params_sampler,
test_task_params_sampler=test_task_params_sampler,
**variant['algo_params']
)
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
return 1
def experiment(variant):
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)
expert_replay_buffer = buffer_save_dict['train']
if 'minmax_env_with_demo_stats' in variant.keys():
if variant['minmax_env_with_demo_stats']:
print('Use minmax envs')
assert 'norm_train' in buffer_save_dict.keys()
expert_replay_buffer = buffer_save_dict['norm_train']
env_specs = variant['env_specs']
env = get_env(env_specs)
env.seed(env_specs['eval_env_seed'])
training_env = get_env(env_specs)
training_env.seed(env_specs['training_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']:
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'],
)
training_env = ScaledEnv(
training_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'],
)
elif variant['minmax_env_with_demo_stats']:
env = MinmaxEnv(
env,
obs_min=buffer_save_dict['obs_min'],
obs_max=buffer_save_dict['obs_max'],
)
training_env = MinmaxEnv(
training_env,
obs_min=buffer_save_dict['obs_min'],
obs_max=buffer_save_dict['obs_max'],
)
obs_space = env.observation_space
act_space = env.action_space
assert not isinstance(obs_space, Dict)
assert len(obs_space.shape) == 1
assert len(act_space.shape) == 1
obs_dim = obs_space.shape[0]
action_dim = act_space.shape[0]
# build the policy models
net_size = variant['policy_net_size']
num_hidden = variant['policy_num_hidden_layers']
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=num_hidden * [net_size],
obs_dim=obs_dim,
action_dim=action_dim,
)
input_dim = obs_dim + action_dim if not variant['ebm_params']['state_only'] else 2*obs_dim
# build the energy model
if (variant['ebm_params']['mode']) == 'deen':
ebm_model = MLPEBM(
input_dim,
num_layer_blocks=variant['ebm_num_blocks'],
hid_dim=variant['ebm_hid_dim'],
hid_act=variant['ebm_hid_act'],
use_bn=variant['ebm_use_bn'],
clamp_magnitude=variant['ebm_clamp_magnitude']
)
algorithm = EBMLearn(
env=env,
training_env=training_env,
ebm=ebm_model,
input_dim = input_dim,
exploration_policy=policy,
sigma=variant['sigma'],
expert_replay_buffer=expert_replay_buffer,
**variant['ebm_params']
)
# build the energy model
elif (variant['ebm_params']['mode']) == 'ae':
ebm_model = MLPAE(
input_dim,
num_layer_blocks=variant['ebm_num_blocks'],
hid_dim=variant['ebm_hid_dim'],
hid_act=variant['ebm_hid_act'],
use_bn=variant['ebm_use_bn'],
)
algorithm = EBMLearn(
env=env,
training_env=training_env,
ebm=ebm_model,
input_dim = input_dim,
exploration_policy=policy,
sigma=None,
expert_replay_buffer=expert_replay_buffer,
**variant['ebm_params']
)
if ptu.gpu_enabled():
algorithm.to(ptu.device)
algorithm.train()
return 1
def experiment(variant):
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)
expert_replay_buffer = buffer_save_dict['train']
if 'minmax_env_with_demo_stats' in variant.keys():
if variant['minmax_env_with_demo_stats']:
print('Use minmax envs')
assert 'norm_train' in buffer_save_dict.keys()
expert_replay_buffer = buffer_save_dict['norm_train']
env_specs = variant['env_specs']
env = get_env(env_specs)
env.seed(env_specs['eval_env_seed'])
training_env = get_env(env_specs)
training_env.seed(env_specs['training_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']:
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'],
)
training_env = ScaledEnv(
training_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'],
)
elif variant['minmax_env_with_demo_stats']:
env = MinmaxEnv(
env,
obs_min=buffer_save_dict['obs_min'],
obs_max=buffer_save_dict['obs_max'],
)
training_env = MinmaxEnv(
training_env,
obs_min=buffer_save_dict['obs_min'],
obs_max=buffer_save_dict['obs_max'],
)
obs_space = env.observation_space
act_space = env.action_space
assert not isinstance(obs_space, Dict)
assert len(obs_space.shape) == 1
assert len(act_space.shape) == 1
obs_dim = obs_space.shape[0]
action_dim = act_space.shape[0]
# build the policy models
net_size = variant['policy_net_size']
num_hidden = variant['policy_num_hidden_layers']
qf1 = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim + action_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim + action_dim,
output_size=1,
)
vf = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=num_hidden * [net_size],
obs_dim=obs_dim,
action_dim=action_dim,
)
# build the energy model
if variant['ebil_params']['mode'] == 'deen':
"""
ebm_model = MLPEBM(
obs_dim + action_dim if not variant['ebil_params']['state_only'] else 2*obs_dim,
num_layer_blocks=variant['ebm_num_blocks'],
hid_dim=variant['ebm_hid_dim'],
hid_act=variant['ebm_hid_act'],
use_bn=variant['ebm_use_bn'],
clamp_magnitude=variant['ebm_clamp_magnitude'],
)
"""
ebm_exp_name = 'ebm-deen-' + variant['env_specs'][
'env_name'] + '-' + str(
variant['expert_traj_num']) + '-train--sigma-' + str(
variant['ebm_sigma'])
ebm_dir = os.path.join(config.LOCAL_LOG_DIR, ebm_exp_name)
ebm_id_dirs = os.listdir(ebm_dir)
ebm_id_dirs = sorted(
ebm_id_dirs,
key=lambda x: os.path.getmtime(os.path.join(ebm_dir, x)))
load_ebm_dir = os.path.join(
ebm_dir, ebm_id_dirs[-1]) # Choose the last as the load ebm dir
load_epoch = variant['ebm_epoch']
load_name = 'itr_{}.pkl'.format(load_epoch)
if load_epoch == 'best':
load_name = 'best.pkl'
load_ebm_path = os.path.join(load_ebm_dir, load_name)
load_ebm_pkl = joblib.load(load_ebm_path, mmap_mode='r')
ebm_model = load_ebm_pkl['ebm']
elif variant['ebil_params']['mode'] == 'ae':
ebm_exp_name = 'ebm-ae-' + variant['env_specs']['env_name'] + '-' + str(
variant['expert_traj_num']) + '-train--sigma-' + str(
variant['ebm_sigma'])
ebm_dir = os.path.join(config.LOCAL_LOG_DIR, ebm_exp_name)
ebm_id_dirs = os.listdir(ebm_dir)
ebm_id_dirs = sorted(
ebm_id_dirs,
key=lambda x: os.path.getmtime(os.path.join(ebm_dir, x)))
load_ebm_dir = os.path.join(
ebm_dir, ebm_id_dirs[-1]) # Choose the last as the load ebm dir
load_epoch = variant['ebm_epoch']
load_name = 'itr_{}.pkl'.format(load_epoch)
if load_epoch == 'best':
load_name = 'best.pkl'
load_ebm_path = os.path.join(load_ebm_dir, load_name)
load_ebm_pkl = joblib.load(load_ebm_path, mmap_mode='r')
ebm_model = load_ebm_pkl['ebm']
else:
raise NotImplementedError
print("loaded EBM from {}".format(load_ebm_path))
# Test
if variant['test']:
batch_data = expert_replay_buffer.random_batch(
100, keys=['observations', 'actions'])
print('ebm_obs: ', np.mean(batch_data['observations'], axis=0))
obs = torch.Tensor(batch_data['observations'])
acts = torch.Tensor(batch_data['actions'])
exp_input = torch.cat([obs, acts], dim=1).to(ptu.device)
print("Not expert data", ebm_model(exp_input * 200).mean().item())
print("Expert data", ebm_model(exp_input).mean().item())
exit(1)
# set up the algorithm
trainer = SoftActorCritic(policy=policy,
qf1=qf1,
qf2=qf2,
vf=vf,
**variant['sac_params'])
algorithm_pretrain = BC(env=env,
training_env=training_env,
exploration_policy=policy,
expert_replay_buffer=expert_replay_buffer,
**variant['bc_params'])
algorithm = EBIL(env=env,
training_env=training_env,
exploration_policy=policy,
rew_func=variant['rew_func'],
cons=variant['cons'],
ebm=ebm_model,
policy_trainer=trainer,
expert_replay_buffer=expert_replay_buffer,
**variant['ebil_params'])
if ptu.gpu_enabled():
algorithm_pretrain.to(ptu.device)
algorithm.to(ptu.device)
else:
algorithm_pretrain.to('cpu')
algorithm.to('cpu')
if variant['pretrain']:
algorithm_pretrain.train()
algorithm.train()
return 1
def experiment(variant):
env_specs = variant['env_specs']
env = get_env(env_specs)
env.seed(env_specs['eval_env_seed'])
training_env = get_env(env_specs)
training_env.seed(env_specs['training_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))
obs_space = env.observation_space
act_space = env.action_space
assert not isinstance(obs_space, Dict)
assert len(obs_space.shape) == 1
assert len(act_space.shape) == 1
obs_dim = obs_space.shape[0]
action_dim = act_space.shape[0]
net_size = variant['net_size']
num_hidden = variant['num_hidden_layers']
qf1 = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim + action_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim + action_dim,
output_size=1,
)
vf = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=num_hidden * [net_size],
obs_dim=obs_dim,
action_dim=action_dim,
)
trainer = SoftActorCritic(
policy=policy,
qf1=qf1,
qf2=qf2,
vf=vf,
**variant['sac_params']
)
algorithm = TorchRLAlgorithm(
trainer=trainer,
env=env,
training_env=training_env,
exploration_policy=policy,
**variant['rl_alg_params']
)
if ptu.gpu_enabled():
algorithm.to(ptu.device)
algorithm.train()
return 1
def experiment(variant):
# get the expert data
with open(EXPERT_LISTING_YAML_PATH, 'r') as f:
listings = yaml.load(f.read())
expert_dir = listings[variant['expert_name']]['exp_dir']
specific_run = listings[variant['expert_name']]['seed_runs'][
variant['expert_seed_run_idx']]
file_to_load = path.join(expert_dir, specific_run, 'extra_data.pkl')
extra_data = joblib.load(file_to_load)
expert_buffer = extra_data['train']
# set up the env
env_specs = variant['env_specs']
if env_specs['train_test_env']:
env, training_env = get_env(env_specs)
else:
env, _ = get_env(env_specs)
training_env, _ = get_env(env_specs)
if variant['scale_env_with_given_demo_stats']:
assert not env_specs['normalized']
env = ScaledEnv(
env,
obs_mean=extra_data['obs_mean'],
obs_std=extra_data['obs_std'],
acts_mean=extra_data['acts_mean'],
acts_std=extra_data['acts_std'],
)
training_env = ScaledEnv(
training_env,
obs_mean=extra_data['obs_mean'],
obs_std=extra_data['obs_std'],
acts_mean=extra_data['acts_mean'],
acts_std=extra_data['acts_std'],
)
# seed the env
env.seed(variant['seed'])
training_env.seed(variant['seed'])
print(env.observation_space)
# compute obs_dim and action_dim
if isinstance(env.observation_space, Dict):
if not variant['algo_params']['policy_uses_pixels']:
obs_dim = int(np.prod(env.observation_space.spaces['obs'].shape))
if variant['algo_params']['policy_uses_task_params']:
if variant['algo_params']['concat_task_params_to_policy_obs']:
obs_dim += int(
np.prod(env.observation_space.
spaces['obs_task_params'].shape))
else:
raise NotImplementedError()
else:
raise NotImplementedError()
else:
obs_dim = int(np.prod(env.observation_space.shape))
action_dim = int(np.prod(env.action_space.shape))
print(obs_dim, action_dim)
sleep(3)
# set up the policy models
policy_net_size = variant['policy_net_size']
hidden_sizes = [policy_net_size] * variant['policy_num_hidden_layers']
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=hidden_sizes,
obs_dim=obs_dim,
action_dim=action_dim,
batch_norm=variant['policy_uses_bn'],
layer_norm=variant['policy_uses_layer_norm'])
# policy = MlpPolicy(
# hidden_sizes=hidden_sizes,
# obs_dim=obs_dim,
# action_dim=action_dim,
# batch_norm=variant['policy_uses_bn'],
# layer_norm=variant['policy_uses_layer_norm']
# )
# set up the AIRL algorithm
algorithm = BC(env,
policy,
expert_buffer,
training_env=training_env,
wrap_absorbing=variant['wrap_absorbing_state'],
**variant['algo_params'])
print(algorithm.exploration_policy)
print(algorithm.eval_policy)
# train
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
return 1
def experiment(variant):
env_specs = variant['env_specs']
if variant['algo_params']['meta']:
env, training_env = get_meta_env(env_specs)
else:
if env_specs['train_test_env']:
env, training_env = get_env(env_specs)
else:
env, _ = get_env(env_specs)
training_env, _ = get_env(env_specs)
if variant['algo_params']['meta']:
train_task_params_sampler, test_task_params_sampler = get_meta_env_params_iters(
env_specs)
print(env.observation_space)
if isinstance(env.observation_space, Dict):
if not variant['algo_params']['policy_uses_pixels']:
obs_dim = int(np.prod(env.observation_space.spaces['obs'].shape))
if variant['algo_params']['policy_uses_task_params']:
obs_dim += int(
np.prod(
env.observation_space.spaces['obs_task_params'].shape))
else:
raise NotImplementedError()
else:
obs_dim = int(np.prod(env.observation_space.shape))
action_dim = int(np.prod(env.action_space.shape))
net_size = variant['net_size']
hidden_sizes = [net_size] * variant['num_hidden_layers']
if variant['use_custom_ant_models']:
assert isinstance(env.observation_space, Dict)
print('CUSTOM ANT WITH LINEAR EMBEDDING OF THE TARGET POSITION')
qf1 = AntRandGoalCustomQFunc(
int(np.prod(
env.observation_space.spaces['obs_task_params'].shape)),
variant['goal_embed_dim'],
hidden_sizes=hidden_sizes,
input_size=int(np.prod(env.observation_space.spaces['obs'].shape))
+ action_dim,
output_size=1,
)
qf2 = AntRandGoalCustomQFunc(
int(np.prod(
env.observation_space.spaces['obs_task_params'].shape)),
variant['goal_embed_dim'],
hidden_sizes=hidden_sizes,
input_size=int(np.prod(env.observation_space.spaces['obs'].shape))
+ action_dim,
output_size=1,
)
vf = AntRandGoalCustomVFunc(
int(np.prod(
env.observation_space.spaces['obs_task_params'].shape)),
variant['goal_embed_dim'],
hidden_sizes=hidden_sizes,
input_size=int(np.prod(env.observation_space.spaces['obs'].shape)),
output_size=1,
)
policy = AntRandGoalCustomReparamTanhMultivariateGaussianPolicy(
int(np.prod(
env.observation_space.spaces['obs_task_params'].shape)),
variant['goal_embed_dim'],
hidden_sizes=hidden_sizes,
obs_dim=int(np.prod(env.observation_space.spaces['obs'].shape)),
action_dim=action_dim,
)
# CUSTOM ANT WITH GATING ACTIVATIONS OF EACH LAYER
# qf1 = AntCustomGatingQFuncV1()
# qf2 = AntCustomGatingQFuncV1()
# vf = AntCustomGatingVFuncV1()
# policy = AntCustomGatingV1ReparamTanhMultivariateGaussianPolicy()
else:
print('Using simple model')
qf1 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + action_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + action_dim,
output_size=1,
)
vf = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=hidden_sizes,
obs_dim=obs_dim,
action_dim=action_dim,
)
if variant['algo_params']['meta']:
algorithm = MetaNewSoftActorCritic(
env=env,
training_env=training_env,
policy=policy,
qf1=qf1,
qf2=qf2,
vf=vf,
train_task_params_sampler=train_task_params_sampler,
test_task_params_sampler=test_task_params_sampler,
true_env_obs_dim=int(
np.prod(env.observation_space.spaces['obs'].shape)),
**variant['algo_params'])
else:
algorithm = NewSoftActorCritic(env=env,
training_env=training_env,
policy=policy,
qf1=qf1,
qf2=qf2,
vf=vf,
**variant['algo_params'])
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
return 1
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']]
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'])
training_env = get_env(env_specs)
training_env.seed(env_specs['training_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))
obs_space = env.observation_space
act_space = env.action_space
assert not isinstance(obs_space, Dict)
assert len(obs_space.shape) == 1
assert len(act_space.shape) == 1
obs_dim = obs_space.shape[0]
action_dim = act_space.shape[0]
# build the policy models
net_size = variant['policy_net_size']
num_hidden = variant['policy_num_hidden_layers']
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=num_hidden * [net_size],
obs_dim=obs_dim,
action_dim=action_dim,
)
input_dim = len(state_indices)
# build the energy model
if (variant['ebm_params']['mode']) == 'deen':
ebm_model = MLPEBM(input_dim,
num_layer_blocks=variant['ebm_num_blocks'],
hid_dim=variant['ebm_hid_dim'],
hid_act=variant['ebm_hid_act'],
use_bn=variant['ebm_use_bn'],
clamp_magnitude=variant['ebm_clamp_magnitude'])
algorithm = EBMLearn(env=env,
training_env=training_env,
ebm=ebm_model,
input_dim=input_dim,
exploration_policy=policy,
sigma=variant['sigma'],
target_state_buffer=target_state_buffer,
state_indices=state_indices,
**variant['ebm_params'])
# build the energy model
elif (variant['ebm_params']['mode']) == 'ae':
ebm_model = MLPAE(
input_dim,
num_layer_blocks=variant['ebm_num_blocks'],
hid_dim=variant['ebm_hid_dim'],
hid_act=variant['ebm_hid_act'],
use_bn=variant['ebm_use_bn'],
)
algorithm = EBMLearn(env=env,
training_env=training_env,
ebm=ebm_model,
input_dim=input_dim,
exploration_policy=policy,
sigma=None,
rescale=variant['rescale'],
target_state_buffer=target_state_buffer,
state_indices=state_indices,
**variant['ebm_params'])
if ptu.gpu_enabled():
algorithm.to(ptu.device)
algorithm.train()
return 1
def experiment(variant):
# env = NormalizedBoxEnv(HalfCheetahEnv())
# env = NormalizedBoxEnv(InvertedPendulumEnv())
# ---------
# env = NormalizedBoxEnv(get_meta_env(variant['env_specs']))
# training_env = NormalizedBoxEnv(get_meta_env(variant['env_specs']))
# env = ReacherEnv()
# training_env = ReacherEnv()
# env = NormalizedBoxEnv(ReacherEnv())
# training_env = NormalizedBoxEnv(ReacherEnv())
# Or for a specific version:
# import gym
# env = NormalizedBoxEnv(gym.make('HalfCheetah-v1'))
# we have to generate the combinations for the env_specs
env_specs = variant['env_specs']
env, _ = get_env(env_specs)
training_env, _ = get_env(env_specs)
print(env.observation_space)
obs_space = env.observation_space
if isinstance(env.observation_space, Dict):
# possible keys: pixel, obs, obs_task_params
if not variant['algo_params']['policy_uses_pixels']:
obs_dim = int(np.prod(obs_space.spaces['obs'].shape))
else:
raise NotImplementedError()
if variant['algo_params']['policy_uses_task_params']:
if variant['algo_params']['concat_task_params_to_policy_obs']:
obs_dim += int(
np.prod(obs_space.spaces['obs_task_params'].shape))
else:
raise NotImplementedError
else:
# OpenAI Gym Env or DMCS Env with only one obs
obs_dim = int(np.prod(env.observation_space.shape))
action_dim = int(np.prod(env.action_space.shape))
# if variant['reload_policy_from'] != '':
# params = joblib.load(variant['reload_policy_from'])
# qf1, qf2, vf, policy = params['qf1'], params['qf2'], params['vf'], params['policy']
# else:
net_size = variant['net_size']
qf1 = FlattenMlp(
hidden_sizes=[net_size, net_size],
input_size=obs_dim + action_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=[net_size, net_size],
input_size=obs_dim + action_dim,
output_size=1,
)
vf = FlattenMlp(
hidden_sizes=[net_size, net_size],
input_size=obs_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=[net_size, net_size],
obs_dim=obs_dim,
action_dim=action_dim,
)
algorithm = NewSoftActorCritic(env=env,
training_env=training_env,
policy=policy,
qf1=qf1,
qf2=qf2,
vf=vf,
**variant['algo_params'])
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
return 1
def experiment(variant):
# set up the envs
env_specs = variant['env_specs']
meta_train_env, meta_test_env = get_meta_env(env_specs)
meta_train_env.seed(variant['seed'])
meta_test_env.seed(variant['seed'])
# set up the policy and training algorithm
if isinstance(meta_train_env.observation_space, Dict):
if variant['algo_params']['policy_uses_pixels']:
raise NotImplementedError(
'Not implemented pixel version of things!')
else:
obs_dim = int(
np.prod(meta_train_env.observation_space.spaces['obs'].shape))
else:
obs_dim = int(np.prod(meta_train_env.observation_space.shape))
action_dim = int(np.prod(meta_train_env.action_space.shape))
print('obs dim: %d' % obs_dim)
print('act dim: %d' % action_dim)
sleep(3)
z_dim = variant['algo_params']['z_dim']
policy_net_size = variant['policy_net_size']
hidden_sizes = [policy_net_size] * variant['num_hidden_layers']
qf1 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + action_dim + z_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + action_dim + z_dim,
output_size=1,
)
vf = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + z_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=hidden_sizes,
obs_dim=obs_dim + z_dim,
action_dim=action_dim,
)
# make the encoder
encoder = TimestepBasedEncoder(
2 * obs_dim + action_dim, #(s,a,s')
variant['algo_params']['r_dim'],
variant['algo_params']['z_dim'],
variant['algo_params']['enc_hid_dim'],
variant['algo_params']['r2z_hid_dim'],
variant['algo_params']['num_enc_layer_blocks'],
hid_act='relu',
use_bn=True,
within_traj_agg=variant['algo_params']['within_traj_agg'],
state_only=False)
train_task_params_sampler, test_task_params_sampler = get_meta_env_params_iters(
env_specs)
algorithm = PEARL(
meta_test_env, # env is the test env, training_env is the training env (following rlkit original setup)
policy,
qf1,
qf2,
vf,
encoder,
# z_dim,
training_env=meta_train_env, # the env used for generating trajectories
train_task_params_sampler=train_task_params_sampler,
test_task_params_sampler=test_task_params_sampler,
**variant['algo_params'])
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
return 1
def experiment(variant):
with open(EXPERT_LISTING_YAML_PATH, 'r') as f:
listings = yaml.load(f.read())
expert_dir = listings[variant['expert_name']]['exp_dir']
specific_run = listings[variant['expert_name']]['seed_runs'][
variant['expert_seed_run_idx']]
file_to_load = path.join(expert_dir, specific_run, 'extra_data.pkl')
extra_data = joblib.load(file_to_load)
# this script is for the non-meta-learning AIRL
train_context_buffer, train_test_buffer = extra_data['meta_train'][
'context'], extra_data['meta_train']['test']
test_context_buffer, test_test_buffer = extra_data['meta_test'][
'context'], extra_data['meta_test']['test']
# set up the envs
env_specs = variant['env_specs']
meta_train_env, meta_test_env = get_meta_env(env_specs)
meta_train_env.seed(variant['seed'])
meta_test_env.seed(variant['seed'])
# set up the policy and training algorithm
if isinstance(meta_train_env.observation_space, Dict):
if variant['algo_params']['policy_uses_pixels']:
raise NotImplementedError(
'Not implemented pixel version of things!')
else:
obs_dim = int(
np.prod(meta_train_env.observation_space.spaces['obs'].shape))
else:
obs_dim = int(np.prod(meta_train_env.observation_space.shape))
action_dim = int(np.prod(meta_train_env.action_space.shape))
print('obs dim: %d' % obs_dim)
print('act dim: %d' % action_dim)
sleep(3)
# make the disc model
if variant['algo_params']['state_only']:
print('\n\nUSING STATE ONLY DISC\n\n')
disc_model = StandardMetaDisc(
2 * obs_dim + action_dim + variant['algo_params']['z_dim']
if not variant['algo_params']['state_only'] else 2 * obs_dim +
variant['algo_params']['z_dim'],
num_layer_blocks=variant['disc_num_blocks'],
hid_dim=variant['disc_hid_dim'],
hid_act=variant['disc_hid_act'],
use_bn=variant['disc_use_bn'],
clamp_magnitude=variant['disc_clamp_magnitude'])
print(disc_model)
print(disc_model.clamp_magnitude)
if variant['algo_params']['use_target_disc']:
target_disc = disc_model.copy()
else:
target_disc = None
print(disc_model)
print(disc_model.clamp_magnitude)
z_dim = variant['algo_params']['z_dim']
policy_net_size = variant['policy_net_size']
hidden_sizes = [policy_net_size] * variant['num_hidden_layers']
qf1 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + action_dim + z_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + action_dim + z_dim,
output_size=1,
)
vf = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + z_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=hidden_sizes,
obs_dim=obs_dim + z_dim,
action_dim=action_dim,
)
policy_optimizer = NewSoftActorCritic(
policy=policy,
qf1=qf1,
qf2=qf2,
vf=vf,
wrap_absorbing=variant['algo_params']['wrap_absorbing'],
**variant['policy_params'])
# make the encoder
encoder = TimestepBasedEncoder(
2 * obs_dim + action_dim, #(s,a,s')
variant['algo_params']['r_dim'],
variant['algo_params']['z_dim'],
variant['algo_params']['enc_hid_dim'],
variant['algo_params']['r2z_hid_dim'],
variant['algo_params']['num_enc_layer_blocks'],
hid_act='relu',
use_bn=True,
)
train_task_params_sampler, test_task_params_sampler = get_meta_env_params_iters(
env_specs)
algorithm = MetaFAIRL(
meta_test_env, # env is the test env, training_env is the training env (following rlkit original setup)
policy,
disc_model,
train_context_buffer,
train_test_buffer,
test_context_buffer,
test_test_buffer,
encoder,
policy_optimizer,
training_env=meta_train_env, # the env used for generating trajectories
train_task_params_sampler=train_task_params_sampler,
test_task_params_sampler=test_task_params_sampler,
target_disc=target_disc,
**variant['algo_params'])
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
return 1
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'])
training_env = get_env(env_specs)
training_env.seed(env_specs['training_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))
obs_space = env.observation_space
act_space = env.action_space
assert not isinstance(obs_space, Dict)
assert len(obs_space.shape) == 1
assert len(act_space.shape) == 1
obs_dim = obs_space.shape[0]
action_dim = act_space.shape[0]
# build the policy models
net_size = variant['policy_net_size']
num_hidden = variant['policy_num_hidden_layers']
qf1 = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim + action_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim + action_dim,
output_size=1,
)
vf = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=num_hidden * [net_size],
obs_dim=obs_dim,
action_dim=action_dim,
)
# build the energy model
if variant['ebil_params']['mode'] == 'deen':
"""
ebm_model = MLPEBM(
obs_dim + action_dim if not variant['ebil_params']['state_only'] else 2*obs_dim,
num_layer_blocks=variant['ebm_num_blocks'],
hid_dim=variant['ebm_hid_dim'],
hid_act=variant['ebm_hid_act'],
use_bn=variant['ebm_use_bn'],
clamp_magnitude=variant['ebm_clamp_magnitude'],
)
"""
ebm_exp_name = 'ebm-deen-' + variant['env_specs'][
'env_name'] + '-' + str(
variant['expert_traj_num']) + '-train--sigma-' + str(
variant['ebm_sigma'])
ebm_dir = os.path.join(config.LOCAL_LOG_DIR, ebm_exp_name)
load_ebm_dir = ebm_dir
load_epoch = variant['ebm_epoch']
load_name = 'itr_{}.pkl'.format(load_epoch)
if load_epoch == 'best':
load_name = 'best.pkl'
load_ebm_path = os.path.join(load_ebm_dir, load_name)
load_ebm_pkl = joblib.load(load_ebm_path, mmap_mode='r')
ebm_model = load_ebm_pkl['ebm']
else:
raise NotImplementedError
# Test
if variant['test']:
batch_data = target_state_buffer / variant['rescale']
obs = torch.Tensor(batch_data[:1000]).to(ptu.device)
print("Not expert data", ebm_model(obs * 200).mean().item())
print("Expert data", ebm_model(obs).mean().item())
exit(1)
# set up the algorithm
trainer = SoftActorCritic(policy=policy,
qf1=qf1,
qf2=qf2,
vf=vf,
**variant['sac_params'])
algorithm = EBIL(env=env,
training_env=training_env,
exploration_policy=policy,
rew_func=variant['rew_func'],
cons=variant['cons'],
rescale=variant['rescale'],
ebm=ebm_model,
policy_trainer=trainer,
target_state_buffer=target_state_buffer,
state_indices=state_indices,
**variant['ebil_params'])
if ptu.gpu_enabled():
algorithm.to(ptu.device)
algorithm.train()
return 1
def experiment(variant):
with open(EXPERT_LISTING_YAML_PATH, 'r') as f:
listings = yaml.load(f.read())
expert_dir = listings[variant['expert_name']]['exp_dir']
specific_run = listings[variant['expert_name']]['seed_runs'][
variant['expert_seed_run_idx']]
file_to_load = path.join(expert_dir, specific_run, 'extra_data.pkl')
extra_data = joblib.load(file_to_load)
# this script is for the non-meta-learning GAIL
expert_buffer = extra_data['train']
# set up the env
env_specs = variant['env_specs']
if env_specs['train_test_env']:
env, training_env = get_env(env_specs)
else:
env, _ = get_env(env_specs)
training_env, _ = get_env(env_specs)
env.seed(variant['seed'])
training_env.seed(variant['seed'])
print(env.observation_space)
if variant['scale_env_with_given_demo_stats']:
assert not env_specs['normalized']
env = ScaledEnv(
env,
obs_mean=extra_data['obs_mean'],
obs_std=extra_data['obs_std'],
acts_mean=extra_data['acts_mean'],
acts_std=extra_data['acts_std'],
)
training_env = ScaledEnv(
training_env,
obs_mean=extra_data['obs_mean'],
obs_std=extra_data['obs_std'],
acts_mean=extra_data['acts_mean'],
acts_std=extra_data['acts_std'],
)
# compute obs_dim and action_dim
if isinstance(env.observation_space, Dict):
if not variant['algo_params']['policy_uses_pixels']:
obs_dim = int(np.prod(env.observation_space.spaces['obs'].shape))
if variant['algo_params']['policy_uses_task_params']:
if variant['algo_params']['concat_task_params_to_policy_obs']:
obs_dim += int(
np.prod(env.observation_space.
spaces['obs_task_params'].shape))
else:
raise NotImplementedError()
else:
raise NotImplementedError()
else:
obs_dim = int(np.prod(env.observation_space.shape))
action_dim = int(np.prod(env.action_space.shape))
print(obs_dim, action_dim)
sleep(3)
# set up the policy models
policy_net_size = variant['policy_net_size']
hidden_sizes = [policy_net_size] * variant['policy_num_hidden_layers']
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=hidden_sizes,
obs_dim=obs_dim,
action_dim=action_dim,
)
# set up the discriminator models
disc_model = StandardAIRLDisc(
obs_dim + action_dim,
num_layer_blocks=variant['disc_num_blocks'],
hid_dim=variant['disc_hid_dim'],
hid_act=variant['disc_hid_act'],
use_bn=variant['disc_use_bn'],
clamp_magnitude=variant['disc_clamp_magnitude'])
print(disc_model)
print(disc_model.clamp_magnitude)
# set up the AdvBC algorithm
algorithm = AdvBC(env,
policy,
disc_model,
expert_buffer,
training_env=training_env,
wrap_absorbing=variant['wrap_absorbing_state'],
**variant['algo_params'])
print(algorithm.use_target_disc)
print(algorithm.soft_target_disc_tau)
print(algorithm.exploration_policy)
print(algorithm.eval_policy)
print(algorithm.disc_optimizer.defaults['lr'])
print(algorithm.policy_optimizer.defaults['lr'])
# train
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
return 1
def experiment(variant):
print('RUNNING')
# we have to generate the combinations for the env_specs
env_specs = variant['env_specs']
env_sampler = OnTheFlyEnvSampler(env_specs)
# Build the normalizer for the env params
env_spec_ranges = {}
for k, v in env_specs.items():
if isinstance(v, list):
env_spec_ranges[k] = v
mean = []
half_diff = []
for k in sorted(env_spec_ranges.keys()):
r = env_spec_ranges[k]
mean.append((r[0] + r[1]) / 2.0)
half_diff.append((r[1] - r[0]) / 2.0)
mean = np.array(mean)
half_diff = np.array(half_diff)
def env_params_normalizer(params):
return (params - mean) / half_diff
variant['algo_params']['env_params_normalizer'] = env_params_normalizer
# set up similar to non-meta version
sample_env, _ = env_sampler()
if variant['algo_params']['concat_env_params_to_obs']:
meta_params_dim = sample_env.env_meta_params.shape[0]
else:
meta_params_dim = 0
obs_dim = int(np.prod(sample_env.observation_space.shape))
action_dim = int(np.prod(sample_env.action_space.shape))
net_size = variant['net_size']
qf1 = FlattenMlp(
hidden_sizes=[net_size, net_size],
input_size=obs_dim + action_dim + meta_params_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=[net_size, net_size],
input_size=obs_dim + action_dim + meta_params_dim,
output_size=1,
)
vf = FlattenMlp(
hidden_sizes=[net_size, net_size],
input_size=obs_dim + meta_params_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=[net_size, net_size],
obs_dim=obs_dim + meta_params_dim,
action_dim=action_dim,
)
algorithm = NewMetaSoftActorCritic(env_sampler=env_sampler,
policy=policy,
qf1=qf1,
qf2=qf2,
vf=vf,
**variant['algo_params'])
# assert False, "Have not added new sac yet!"
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
return 1
def experiment(variant):
expert_buffer = joblib.load(variant['xy_data_path'])['xy_data']
# set up the env
env_specs = variant['env_specs']
if env_specs['train_test_env']:
env, training_env = get_env(env_specs)
else:
env, _ = get_env(env_specs)
training_env, _ = get_env(env_specs)
env.seed(variant['seed'])
training_env.seed(variant['seed'])
print(env.observation_space)
if variant['scale_env_with_given_demo_stats']:
assert False
assert not env_specs['normalized']
env = ScaledEnv(
env,
obs_mean=extra_data['obs_mean'],
obs_std=extra_data['obs_std'],
acts_mean=extra_data['acts_mean'],
acts_std=extra_data['acts_std'],
)
training_env = ScaledEnv(
training_env,
obs_mean=extra_data['obs_mean'],
obs_std=extra_data['obs_std'],
acts_mean=extra_data['acts_mean'],
acts_std=extra_data['acts_std'],
)
# compute obs_dim and action_dim
if isinstance(env.observation_space, Dict):
if not variant['algo_params']['policy_uses_pixels']:
obs_dim = int(np.prod(env.observation_space.spaces['obs'].shape))
if variant['algo_params']['policy_uses_task_params']:
if variant['algo_params']['concat_task_params_to_policy_obs']:
obs_dim += int(
np.prod(env.observation_space.
spaces['obs_task_params'].shape))
else:
raise NotImplementedError()
else:
raise NotImplementedError()
else:
obs_dim = int(np.prod(env.observation_space.shape))
action_dim = int(np.prod(env.action_space.shape))
print(obs_dim, action_dim)
sleep(3)
# set up the policy models
policy_net_size = variant['policy_net_size']
hidden_sizes = [policy_net_size] * variant['policy_num_hidden_layers']
qf1 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + action_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + action_dim,
output_size=1,
)
target_qf1 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + action_dim,
output_size=1,
)
target_qf2 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + action_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
# policy = ReparamMultivariateGaussianPolicy(
hidden_sizes=hidden_sizes,
obs_dim=obs_dim,
action_dim=action_dim,
# std=0.1
)
# set up the discriminator models
disc_model_class = ThreeWayResNetAIRLDisc if variant[
'threeway'] else ResNetAIRLDisc
disc_model = disc_model_class(
2, # obs is just x-y pos
num_layer_blocks=variant['disc_num_blocks'],
hid_dim=variant['disc_hid_dim'],
hid_act=variant['disc_hid_act'],
use_bn=variant['disc_use_bn'],
clamp_magnitude=variant['disc_clamp_magnitude'])
print(disc_model)
print(disc_model.clamp_magnitude)
# set up the RL algorithm used to train the policy
policy_optimizer = EntConstSAC(policy=policy,
qf1=qf1,
qf2=qf2,
target_qf1=target_qf1,
target_qf2=target_qf2,
action_dim=action_dim,
**variant['policy_params'])
# set up the AIRL algorithm
alg_class = ThreewayStateMarginalMatchingAlg if variant[
'threeway'] else StateMarginalMatchingAlg
algorithm = alg_class(env,
policy,
disc_model,
policy_optimizer,
expert_buffer,
training_env=training_env,
**variant['algo_params'])
print(algorithm.exploration_policy)
print(algorithm.eval_policy)
print(algorithm.policy_optimizer.policy_optimizer.defaults['lr'])
print(algorithm.policy_optimizer.qf1_optimizer.defaults['lr'])
print(algorithm.policy_optimizer.qf2_optimizer.defaults['lr'])
print(algorithm.disc_optimizer.defaults['lr'])
# train
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
return 1
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']]
buffer_save_dict = joblib.load(demos_path)
target_state_buffer = buffer_save_dict['data']
state_indices = torch.LongTensor(variant['state_indices'])
env_specs = variant['env_specs']
env = get_env(env_specs)
env.seed(env_specs['eval_env_seed'])
training_env = get_env(env_specs)
training_env.seed(env_specs['training_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))
obs_space = env.observation_space
act_space = env.action_space
assert not isinstance(obs_space, Dict)
assert len(obs_space.shape) == 1
assert len(act_space.shape) == 1
obs_dim = obs_space.shape[0]
action_dim = act_space.shape[0]
# build the policy models
net_size = variant['policy_net_size']
num_hidden = variant['policy_num_hidden_layers']
qf1 = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim + action_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim + action_dim,
output_size=1,
)
vf = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=num_hidden * [net_size],
obs_dim=obs_dim,
action_dim=action_dim,
)
# build the discriminator model
if variant['disc_model_type'] == 'resnet_disc':
disc_model = ResNetAIRLDisc(
len(state_indices),
num_layer_blocks=variant['disc_num_blocks'],
hid_dim=variant['disc_hid_dim'],
hid_act=variant['disc_hid_act'],
use_bn=variant['disc_use_bn'],
clamp_magnitude=variant['disc_clamp_magnitude'])
else:
disc_model = MLPDisc(len(state_indices),
num_layer_blocks=variant['disc_num_blocks'],
hid_dim=variant['disc_hid_dim'],
hid_act=variant['disc_hid_act'],
use_bn=variant['disc_use_bn'],
clamp_magnitude=variant['disc_clamp_magnitude'])
# set up the algorithm
trainer = SoftActorCritic(policy=policy,
qf1=qf1,
qf2=qf2,
vf=vf,
**variant['sac_params'])
algorithm = AdvSMM(env=env,
training_env=training_env,
exploration_policy=policy,
discriminator=disc_model,
policy_trainer=trainer,
target_state_buffer=target_state_buffer,
state_indices=state_indices,
**variant['adv_smm_params'])
if ptu.gpu_enabled():
algorithm.to(ptu.device)
algorithm.train()
return 1
def experiment(variant):
with open(EXPERT_LISTING_YAML_PATH, 'r') as f:
listings = yaml.load(f.read())
expert_dir = listings[variant['expert_name']]['exp_dir']
specific_run = listings[variant['expert_name']]['seed_runs'][variant['expert_seed_run_idx']]
file_to_load = path.join(expert_dir, specific_run, 'extra_data.pkl')
extra_data = joblib.load(file_to_load)
# this script is for the non-meta-learning GAIL
train_context_buffer, train_test_buffer = extra_data['meta_train']['context'], extra_data['meta_train']['test']
test_context_buffer, test_test_buffer = extra_data['meta_test']['context'], extra_data['meta_test']['test']
# set up the envs
env_specs = variant['env_specs']
meta_train_env, meta_test_env = get_meta_env(env_specs)
# set up the policy and training algorithm
if isinstance(meta_train_env.observation_space, Dict):
if variant['algo_params']['policy_uses_pixels']:
raise NotImplementedError('Not implemented pixel version of things!')
else:
obs_dim = int(np.prod(meta_train_env.observation_space.spaces['obs'].shape))
else:
obs_dim = int(np.prod(meta_train_env.observation_space.shape))
action_dim = int(np.prod(meta_train_env.action_space.shape))
print('obs dim: %d' % obs_dim)
print('act dim: %d' % action_dim)
sleep(3)
# make the policy and policy optimizer
hidden_sizes = [variant['algo_params']['policy_net_size']] * variant['algo_params']['policy_num_layers']
z_dim = variant['algo_params']['np_params']['z_dim']
qf1 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + z_dim + action_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + z_dim + action_dim,
output_size=1,
)
vf = FlattenMlp(
hidden_sizes=hidden_sizes,
input_size=obs_dim + z_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=hidden_sizes,
obs_dim=obs_dim + z_dim,
action_dim=action_dim,
)
# disc_model = GAILDiscModel(obs_dim + action_dim + z_dim, hid_dim=variant['algo_params']['disc_net_size'])
disc_model = MlpGAILDisc(
hidden_sizes=variant['disc_hidden_sizes'],
output_size=1,
input_size=obs_dim + action_dim + z_dim,
hidden_activation=torch.nn.functional.tanh,
layer_norm=variant['disc_uses_layer_norm']
# output_activation=identity,
)
policy_optimizer = NewSoftActorCritic(
policy=policy,
qf1=qf1,
qf2=qf2,
vf=vf,
**variant['algo_params']['policy_params']
)
# Make the neural process
# in the initial version we are assuming all trajectories have the same length
timestep_enc_params = variant['algo_params']['np_params']['traj_enc_params']['timestep_enc_params']
traj_enc_params = variant['algo_params']['np_params']['traj_enc_params']['traj_enc_params']
timestep_enc_params['input_size'] = obs_dim + action_dim
traj_samples, _ = train_context_buffer.sample_trajs(1, num_tasks=1)
# len_context_traj = traj_samples[0][0]['observations'].shape[0]
len_context_traj = 5
traj_enc_params['input_size'] = timestep_enc_params['output_size'] * len_context_traj
traj_enc = TrivialTrajEncoder(
timestep_enc_params,
traj_enc_params
)
trunk_params = variant['algo_params']['np_params']['r2z_map_params']['trunk_params']
trunk_params['input_size'] = traj_enc.output_size
split_params = variant['algo_params']['np_params']['r2z_map_params']['split_heads_params']
split_params['input_size'] = trunk_params['output_size']
split_params['output_size'] = variant['algo_params']['np_params']['z_dim']
r2z_map = TrivialR2ZMap(
trunk_params,
split_params
)
np_enc = TrivialNPEncoder(
variant['algo_params']['np_params']['np_enc_params']['agg_type'],
traj_enc,
r2z_map
)
# class StupidDistFormat():
# def __init__(self, var):
# self.mean = var
# class ZeroModule(nn.Module):
# def __init__(self, z_dim):
# super().__init__()
# self.z_dim = z_dim
# self.fc = nn.Linear(10,10)
# def forward(self, context):
# c_len = len(context)
# return StupidDistFormat(Variable(torch.zeros(c_len, self.z_dim), requires_grad=False))
# np_enc = ZeroModule(variant['algo_params']['np_params']['z_dim'])
train_task_params_sampler, test_task_params_sampler = get_meta_env_params_iters(env_specs)
algorithm = NeuralProcessAIRL(
meta_test_env, # env is the test env, training_env is the training env (following rlkit original setup)
policy,
disc_model,
train_context_buffer,
train_test_buffer,
test_context_buffer,
test_test_buffer,
np_enc,
policy_optimizer,
training_env=meta_train_env, # the env used for generating trajectories
train_task_params_sampler=train_task_params_sampler,
test_task_params_sampler=test_task_params_sampler,
**variant['algo_params']
)
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
return 1
def experiment(variant):
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)
expert_replay_buffer = buffer_save_dict['train']
env_specs = variant['env_specs']
env = get_env(env_specs)
env.seed(env_specs['eval_env_seed'])
training_env = get_env(env_specs)
training_env.seed(env_specs['training_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']:
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'],
)
training_env = ScaledEnv(
training_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'],
)
obs_space = env.observation_space
act_space = env.action_space
assert not isinstance(obs_space, Dict)
assert len(obs_space.shape) == 1
assert len(act_space.shape) == 1
obs_dim = obs_space.shape[0]
action_dim = act_space.shape[0]
# build the policy models
net_size = variant['policy_net_size']
num_hidden = variant['policy_num_hidden_layers']
qf1 = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim + action_dim,
output_size=1,
)
qf2 = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim + action_dim,
output_size=1,
)
vf = FlattenMlp(
hidden_sizes=num_hidden * [net_size],
input_size=obs_dim,
output_size=1,
)
policy = ReparamTanhMultivariateGaussianPolicy(
hidden_sizes=num_hidden * [net_size],
obs_dim=obs_dim,
action_dim=action_dim,
)
# build the discriminator model
disc_model = MLPDisc(
obs_dim +
action_dim if not variant['adv_irl_params']['state_only'] else 2 *
obs_dim,
num_layer_blocks=variant['disc_num_blocks'],
hid_dim=variant['disc_hid_dim'],
hid_act=variant['disc_hid_act'],
use_bn=variant['disc_use_bn'],
clamp_magnitude=variant['disc_clamp_magnitude'])
# set up the algorithm
trainer = SoftActorCritic(policy=policy,
qf1=qf1,
qf2=qf2,
vf=vf,
**variant['sac_params'])
algorithm = AdvIRL(env=env,
training_env=training_env,
exploration_policy=policy,
discriminator=disc_model,
policy_trainer=trainer,
expert_replay_buffer=expert_replay_buffer,
**variant['adv_irl_params'])
if ptu.gpu_enabled():
algorithm.to(ptu.device)
algorithm.train()
return 1
