def main(config):
set_seed(config['seed'])
baseline = globals()[config['baseline']]() #instantiate baseline
env = globals()[config['env']]() # instantiate env
env = normalize(env) # apply normalize wrapper to env
print("env.observation_space.shape: ", env.observation_space.shape)
policy = MetaGaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
meta_batch_size=config['meta_batch_size'],
hidden_sizes=config['hidden_sizes'],
)
sampler = MetaSampler(
env=('multiworld', config['env'], env),
policy=policy,
rollouts_per_meta_task=config[
'rollouts_per_meta_task'], # This batch_size is confusing
meta_batch_size=config['meta_batch_size'],
max_path_length=config['max_path_length'],
parallel=config['parallel'],
)
sampler.set_env(env)
sample_processor = MetaSampleProcessor(
baseline=baseline,
discount=config['discount'],
gae_lambda=config['gae_lambda'],
normalize_adv=config['normalize_adv'],
)
algo = ProMP(
policy=policy,
inner_lr=config['inner_lr'],
meta_batch_size=config['meta_batch_size'],
num_inner_grad_steps=config['num_inner_grad_steps'],
learning_rate=config['learning_rate'],
num_ppo_steps=config['num_promp_steps'],
clip_eps=config['clip_eps'],
target_inner_step=config['target_inner_step'],
init_inner_kl_penalty=config['init_inner_kl_penalty'],
adaptive_inner_kl_penalty=config['adaptive_inner_kl_penalty'],
)
trainer = Trainer(
algo=algo,
policy=policy,
env=env,
sampler=sampler,
sample_processor=sample_processor,
n_itr=config['n_itr'],
num_inner_grad_steps=config['num_inner_grad_steps'],
)
trainer.train()
class TestLinearFeatureBaseline(unittest.TestCase):
def setUp(self):
self.random_env = RandomEnv()
self.random_policy = RandomPolicy(1, 1)
self.meta_batch_size = 2
self.batch_size = 10
self.path_length = 100
self.linear = LinearFeatureBaseline()
self.sampler = MetaSampler(self.random_env,
self.random_policy,
self.batch_size,
self.meta_batch_size,
self.path_length,
parallel=True)
def testFit(self):
paths = self.sampler.obtain_samples()
for task in paths.values():
unfit_error = 0
for path in task:
path["returns"] = utils.discount_cumsum(path["rewards"], 0.99)
unfit_pred = self.linear.predict(path)
unfit_error += sum([
np.square(pred - actual)
for pred, actual in zip(unfit_pred, path['returns'])
])
self.linear.fit(task)
fit_error = 0
for path in task:
fit_pred = self.linear.predict(path)
fit_error += sum([
np.square(pred - actual)
for pred, actual in zip(fit_pred, path['returns'])
])
self.assertTrue(fit_error < unfit_error)
def testSerialize(self):
paths = self.sampler.obtain_samples()
for task in paths.values():
for path in task:
path["returns"] = utils.discount_cumsum(path["rewards"], 0.99)
self.linear.fit(task)
fit_error_pre = 0
for path in task:
fit_pred = self.linear.predict(path)
fit_error_pre += sum([
np.square(pred - actual)
for pred, actual in zip(fit_pred, path['returns'])
])
pkl = pickle.dumps(self.linear)
self.linear = pickle.loads(pkl)
fit_error_post = 0
for path in task:
fit_pred = self.linear.predict(path)
fit_error_post += sum([
np.square(pred - actual)
for pred, actual in zip(fit_pred, path['returns'])
])
self.assertEqual(fit_error_pre, fit_error_post)
def setUp(self):
self.random_env = RandomEnv()
self.random_policy = RandomPolicy(1, 1)
self.meta_batch_size = 2
self.batch_size = 10
self.path_length = 100
self.linear = LinearFeatureBaseline()
self.sampler = MetaSampler(self.random_env,
self.random_policy,
self.batch_size,
self.meta_batch_size,
self.path_length,
parallel=True)
def main(config):
set_seed(config['seed'])
reward_baseline = LinearTimeBaseline() # the usual baseline
return_baseline = LinearFeatureBaseline(
) # the additional baseline for DICE
env = globals()[config['env']]() # instantiate env
env = normalize(env) # apply normalize wrapper to env
meta_baseline = MetaNNBaseline(
input_size=env.observation_space.shape[0]) # the meta baseline
policy = MetaGaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
meta_batch_size=config['meta_batch_size'],
hidden_sizes=config['hidden_sizes'],
)
sampler = MetaSampler(
env=env,
policy=policy,
rollouts_per_meta_task=config['rollouts_per_meta_task'],
meta_batch_size=config['meta_batch_size'],
max_path_length=config['max_path_length'],
parallel=config['parallel'],
)
sample_processor = TMAMLMetaSampleProcessor(
baseline=reward_baseline,
max_path_length=config['max_path_length'],
discount=config['discount'],
normalize_adv=config['normalize_adv'],
positive_adv=config['positive_adv'],
return_baseline=return_baseline,
metabaseline=meta_baseline,
)
algo = TMAML(policy=policy,
max_path_length=config['max_path_length'],
meta_batch_size=config['meta_batch_size'],
num_inner_grad_steps=config['num_inner_grad_steps'],
inner_lr=config['inner_lr'],
learning_rate=config['learning_rate'])
trainer = Trainer(
algo=algo,
policy=policy,
env=env,
sampler=sampler,
sample_processor=sample_processor,
n_itr=config['n_itr'],
num_inner_grad_steps=config['num_inner_grad_steps'],
)
trainer.train()
def setUp(self):
self.env = env = MetaPointEnv()
self.baseline = baseline = LinearFeatureBaseline()
self.policy = policy = MetaGaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
meta_batch_size=10,
hidden_sizes=(16, 16),
learn_std=True,
hidden_nonlinearity=tf.tanh,
output_nonlinearity=None,
)
self.sampler = MetaSampler(
env=env,
policy=policy,
rollouts_per_meta_task=2,
meta_batch_size=10,
max_path_length=50,
parallel=False,
)
self.sample_processor = MetaSampleProcessor(
baseline=baseline,
discount=0.99,
gae_lambda=1.0,
normalize_adv=True,
positive_adv=False,
)
self.algo = ProMP(
policy=policy,
inner_lr=0.1,
meta_batch_size=10,
num_inner_grad_steps=2,
learning_rate=1e-3,
num_ppo_steps=5,
num_minibatches=1,
clip_eps=0.5,
target_inner_step=2e-2,
init_inner_kl_penalty=1e-3,
)
def main(config):
# config['seed'] = 4
experiment.set_name("short meta saving test")
set_seed(config['seed'])
experiment.log_parameters(config)
# experiment.log_parameter("task limit size", 3)
# experiment.log_metric("seed", config['seed'])
baseline = globals()[config['baseline']]() #instantiate baseline
env = globals()[config['env']]() # instantiate env
env = normalize(env) # apply normalize wrapper to env
with open('/saved_policies/mjvel.policy', 'rb') as policy_file:
policy = pickle.load(policy_file)
print("policy loaded")
sampler = MetaSampler(
env=env,
policy=policy,
rollouts_per_meta_task=config[
'rollouts_per_meta_task'], # This batch_size is confusing
meta_batch_size=config['meta_batch_size'],
max_path_length=config['max_path_length'],
parallel=config['parallel'],
)
sample_processor = MetaSampleProcessor(
baseline=baseline,
discount=config['discount'],
gae_lambda=config['gae_lambda'],
normalize_adv=config['normalize_adv'],
)
algo = ProMP(
policy=policy,
inner_lr=config['inner_lr'],
meta_batch_size=config['meta_batch_size'],
num_inner_grad_steps=config['num_inner_grad_steps'],
learning_rate=config['learning_rate'],
num_ppo_steps=config['num_promp_steps'],
clip_eps=config['clip_eps'],
target_inner_step=config['target_inner_step'],
init_inner_kl_penalty=config['init_inner_kl_penalty'],
adaptive_inner_kl_penalty=config['adaptive_inner_kl_penalty'],
)
trainer = RLTrainer(algo=algo,
policy=policy,
env=env,
sampler=sampler,
sample_processor=sample_processor,
n_itr=config['n_itr'],
num_inner_grad_steps=config['num_inner_grad_steps'],
experiment=experiment)
trainer.train()
def main(config):
set_seed(config['seed'])
sess = tf.Session()
with sess.as_default() as sess:
data = joblib.load(load_path + "/params.pkl")
policy = data['policy']
env = data['env']
baseline = data['baseline']
# config['meta_batch_size'] = env.NUM_EVAL
# policy.meta_batch_size = env.NUM_EVAL
sampler = MetaSampler(
env=env,
policy=policy,
rollouts_per_meta_task=config[
'rollouts_per_meta_task'], # Will be modified later
meta_batch_size=config['meta_batch_size'],
max_path_length=config['max_path_length'],
parallel=config['parallel'],
)
sample_processor = MetaSampleProcessor(
baseline=baseline,
discount=config['discount'],
gae_lambda=config['gae_lambda'],
normalize_adv=config['normalize_adv'],
)
algo = ProMP(
policy=policy,
inner_lr=config['inner_lr'],
meta_batch_size=config['meta_batch_size'],
num_inner_grad_steps=config['num_inner_grad_steps'],
learning_rate=config['learning_rate'],
num_ppo_steps=config['num_promp_steps'],
clip_eps=config['clip_eps'],
target_inner_step=config['target_inner_step'],
init_inner_kl_penalty=config['init_inner_kl_penalty'],
adaptive_inner_kl_penalty=config['adaptive_inner_kl_penalty'],
)
tester = Tester(
algo=algo,
policy=policy,
env=env,
sampler=sampler,
sample_processor=sample_processor,
#n_itr=config['n_itr'],
eff=config['eff'],
num_inner_grad_steps=config['num_inner_grad_steps'],
)
tester.train()
sess.close()
def main(config):
set_seed(config['seed'])
baseline = globals()[config['baseline']]() #instantiate baseline
env = globals()[config['env']]() # instantiate env
env = normalize(env) # apply normalize wrapper to env
policy = MetaGaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
meta_batch_size=config['meta_batch_size'],
hidden_sizes=config['hidden_sizes'],
)
sampler = MetaSampler(
env=env,
policy=policy,
rollouts_per_meta_task=config[
'rollouts_per_meta_task'], # This batch_size is confusing
meta_batch_size=config['meta_batch_size'],
max_path_length=config['max_path_length'],
parallel=config['parallel'],
)
sample_processor = MetaSampleProcessor(
baseline=baseline,
discount=config['discount'],
gae_lambda=config['gae_lambda'],
normalize_adv=config['normalize_adv'],
)
algo = VPGMAML(
policy=policy,
learning_rate=config['learning_rate'],
inner_type=config['inner_type'],
inner_lr=config['inner_lr'],
meta_batch_size=config['meta_batch_size'],
num_inner_grad_steps=config['num_inner_grad_steps'],
exploration=False,
)
trainer = Trainer(
algo=algo,
policy=policy,
env=env,
sampler=sampler,
sample_processor=sample_processor,
n_itr=config['n_itr'],
num_inner_grad_steps=config['num_inner_grad_steps'],
)
trainer.train()
def main(config):
set_seed(config['seed'])
baseline = globals()[config['baseline']]() #instantiate baseline
env = globals()[config['env']]() # instantiate env
env = normalize(env) # apply normalize wrapper to env
policy = MetaGaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
meta_batch_size=config['meta_batch_size'],
hidden_sizes=config['hidden_sizes'],
)
sampler = MetaSampler(
env=env,
policy=policy,
rollouts_per_meta_task=config[
'rollouts_per_meta_task'], # This batch_size is confusing
meta_batch_size=config['meta_batch_size'],
max_path_length=config['max_path_length'],
parallel=config['parallel'],
)
sample_processor = MetaSampleProcessor(
baseline=baseline,
discount=config['discount'],
gae_lambda=config['gae_lambda'],
normalize_adv=config['normalize_adv'],
)
algo = ProMP(
policy=policy,
inner_lr=config['inner_lr'],
meta_batch_size=config['meta_batch_size'],
num_inner_grad_steps=config['num_inner_grad_steps'],
learning_rate=config['learning_rate'],
num_ppo_steps=config['num_promp_steps'],
clip_eps=config['clip_eps'],
target_inner_step=config['target_inner_step'],
init_inner_kl_penalty=config['init_inner_kl_penalty'],
adaptive_inner_kl_penalty=config['adaptive_inner_kl_penalty'],
)
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True # pylint: disable=E1101
sess = tf.Session(config=gpu_config)
saver = tf.train.Saver(
keep_checkpoint_every_n_hours=config['keep_checkpoint_every_n_hours'],
max_to_keep=config['max_checkpoints_to_keep'])
save_path = os.path.join(args.dump_path, 'model.ckpt')
if config['restore_path'] is not None:
logger.log('Restoring parameters from {}'.format(
config['restore_path']))
saver.restore(sess, config['restore_path'])
logger.log('Restored')
trainer = Trainer(
algo=algo,
policy=policy,
env=env,
sampler=sampler,
sample_processor=sample_processor,
saver=saver,
save_path=save_path,
save_steps=config['save_steps'],
n_itr=config['n_itr'],
num_inner_grad_steps=config['num_inner_grad_steps'],
sess=sess,
)
trainer.train()
def main(config):
# config['seed'] = 4ß
# experiment.set_name("pos task only, size = 15, logging vel")
set_seed(config['seed'])
# experiment.log_parameters(config)
# experiment.log_parameter("task limit size", 3)
# experiment.log_metric("seed", config['seed'])
baseline = globals()[config['baseline']]() #instantiate baseline
env = globals()[config['env']]() # instantiate env
print("env: ", env.sample_tasks)
TASKSL1 = np.array([0, -0.3])
env.set_tasks(TASKSL1)
env = normalize(env) # apply normalize wrapper to env
policy = MetaGaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
meta_batch_size=config['meta_batch_size'],
hidden_sizes=config['hidden_sizes'],
)
sampler = MetaSampler(
env=env,
policy=policy,
rollouts_per_meta_task=config[
'rollouts_per_meta_task'], # This batch_size is confusing
meta_batch_size=config['meta_batch_size'],
max_path_length=config['max_path_length'],
parallel=config['parallel'],
)
sample_processor = MetaSampleProcessor(
baseline=baseline,
discount=config['discount'],
gae_lambda=config['gae_lambda'],
normalize_adv=config['normalize_adv'],
)
algo = ProMP(
policy=policy,
inner_lr=config['inner_lr'],
meta_batch_size=config['meta_batch_size'],
num_inner_grad_steps=config['num_inner_grad_steps'],
learning_rate=config['learning_rate'],
num_ppo_steps=config['num_promp_steps'],
clip_eps=config['clip_eps'],
target_inner_step=config['target_inner_step'],
init_inner_kl_penalty=config['init_inner_kl_penalty'],
adaptive_inner_kl_penalty=config['adaptive_inner_kl_penalty'],
)
trainer = Trainer(
algo=algo,
policy=policy,
env=env,
sampler=sampler,
sample_processor=sample_processor,
n_itr=config['n_itr'],
num_inner_grad_steps=config['num_inner_grad_steps'],
)
trainer.train()
def main(args=None):
idx = int(time.time())
args = parse_args(args)
config = {
'seed': args.seed,
'baseline': 'LinearFeatureBaseline',
'env': 'ReachWorld', # not used
'rollouts_per_meta_task': args.rollout_per_meta_task,
'max_path_length': args.max_path_length, # 100
'parallel': not args.seq,
'discount': args.discount,
'gae_lambda': args.gae_lambda,
'normalize_adv': True,
'hidden_sizes': args.hidden_sizes,
'inner_lr': args.inner_lr, # adaptation step size
'learning_rate': args.learning_rate, # meta-policy gradient step size
'num_promp_steps':
args.num_promp_steps, # number of ProMp steps without re-sampling
'clip_eps': args.clip_eps, # clipping range
'target_inner_step': args.target_inner_step,
'init_inner_kl_penalty': args.init_inner_kl_penalty,
'adaptive_inner_kl_penalty': args.
adaptive_inner_kl_penalty, # whether to use an adaptive or fixed KL-penalty coefficient
'n_itr': args.n_itr, # number of overall training iterations
'meta_batch_size':
args.meta_batch_size, # number of sampled meta-tasks per iterations
'num_inner_grad_steps': args.
num_inner_grad_steps, # number of inner / adaptation gradient steps
}
# configure logger
logger.configure(dir=args.dump_path,
format_strs=['stdout', 'log', 'csv'],
snapshot_mode='last_gap')
# dump run configuration before starting training
json.dump(config,
open(args.dump_path + '/params.json', 'w'),
cls=ClassEncoder)
set_seed(config['seed'])
baseline = globals()[config['baseline']]() #instantiate baseline
env = get_env()
#env = normalize(env) # apply normalize wrapper to env
if isinstance(env.action_space, gym.spaces.Box):
action_dim = np.prod(env.action_space.shape)
elif isinstance(env.action_space, gym.spaces.Discrete):
action_dim = env.action_space.n
else:
raise Exception('unknown action space, cannot get action dim')
policy = MetaCategoricalMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=action_dim,
meta_batch_size=config['meta_batch_size'],
hidden_sizes=config['hidden_sizes'],
)
sampler = MetaSampler(
env=env,
policy=policy,
rollouts_per_meta_task=config[
'rollouts_per_meta_task'], # This batch_size is confusing
meta_batch_size=config['meta_batch_size'],
max_path_length=config['max_path_length'],
parallel=config['parallel'],
)
sample_processor = MetaSampleProcessor(
baseline=baseline,
discount=config['discount'],
gae_lambda=config['gae_lambda'],
normalize_adv=config['normalize_adv'],
)
algo = ProMP(
policy=policy,
inner_lr=config['inner_lr'],
meta_batch_size=config['meta_batch_size'],
num_inner_grad_steps=config['num_inner_grad_steps'],
learning_rate=config['learning_rate'],
num_ppo_steps=config['num_promp_steps'],
clip_eps=config['clip_eps'],
target_inner_step=config['target_inner_step'],
init_inner_kl_penalty=config['init_inner_kl_penalty'],
adaptive_inner_kl_penalty=config['adaptive_inner_kl_penalty'],
)
trainer = Trainer(
algo=algo,
policy=policy,
env=env,
sampler=sampler,
sample_processor=sample_processor,
n_itr=config['n_itr'],
num_inner_grad_steps=config['num_inner_grad_steps'],
)
trainer.train()
class TestLikelihoodRation(unittest.TestCase):
"""
Assure that likelihhood ratio at first gradient step is approx. one since pi_old = pi_new
"""
def setUp(self):
self.env = env = MetaPointEnv()
self.baseline = baseline = LinearFeatureBaseline()
self.policy = policy = MetaGaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
meta_batch_size=10,
hidden_sizes=(16, 16),
learn_std=True,
hidden_nonlinearity=tf.tanh,
output_nonlinearity=None,
)
self.sampler = MetaSampler(
env=env,
policy=policy,
rollouts_per_meta_task=2,
meta_batch_size=10,
max_path_length=50,
parallel=False,
)
self.sample_processor = MetaSampleProcessor(
baseline=baseline,
discount=0.99,
gae_lambda=1.0,
normalize_adv=True,
positive_adv=False,
)
self.algo = ProMP(
policy=policy,
inner_lr=0.1,
meta_batch_size=10,
num_inner_grad_steps=2,
learning_rate=1e-3,
num_ppo_steps=5,
num_minibatches=1,
clip_eps=0.5,
target_inner_step=2e-2,
init_inner_kl_penalty=1e-3,
)
def test_likelihood_ratio(self):
with tf.Session() as sess:
# initialize uninitialized vars (only initialize vars that were not loaded)
uninit_vars = [
var for var in tf.global_variables()
if not sess.run(tf.is_variable_initialized(var))
]
sess.run(tf.variables_initializer(uninit_vars))
self.sampler.update_tasks()
self.policy.switch_to_pre_update() # Switch to pre-update policy
all_samples_data, all_paths = [], []
for step in range(1):
""" -------------------- Sampling --------------------------"""
paths = self.sampler.obtain_samples(log_prefix=str(step))
all_paths.append(paths)
""" ----------------- Processing Samples ---------------------"""
samples_data = self.sample_processor.process_samples(paths,
log=False)
all_samples_data.append(samples_data)
""" ------------------- Inner Policy Update --------------------"""
obs_phs, action_phs, adv_phs, dist_info_phs, all_phs = self.algo._make_input_placeholders(
'')
for i in range(self.algo.meta_batch_size):
obs = samples_data[i]['observations']
actions = samples_data[i]['actions']
agent_infos = samples_data[i]['agent_infos']
param_vals = self.policy.get_param_values()
likelihood_ratio_sym = self.policy.likelihood_ratio_sym(
obs_phs[i], action_phs[i], dist_info_phs[i],
self.policy.policies_params_phs[i])
feed_dict_params = dict(
zip(self.policy.policies_params_phs[i].values(),
param_vals.values()))
feed_dict_dist_infos = dict(
zip(dist_info_phs[i].values(), agent_infos.values()))
feed_dict = {obs_phs[i]: obs, action_phs[i]: actions}
feed_dict.update(feed_dict_params)
feed_dict.update(feed_dict_dist_infos)
lr = sess.run(likelihood_ratio_sym, feed_dict=feed_dict)
self.assertTrue(np.allclose(lr, 1))
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
meta_batch_size=params['meta_batch_size'],
hidden_sizes=params['hidden_sizes'],
cell_size=params['cell_size'],
rollouts_per_meta_task=params['rollouts_per_meta_task'],
max_path_length=params['max_path_length'],
use_betas=params['use_betas'],
shift_gammas=params['shift_gammas'],
)
sampler = MetaSampler(
env=env,
policy=policy,
# This batch_size is confusing
rollouts_per_meta_task=params['rollouts_per_meta_task'],
meta_batch_size=params['meta_batch_size'],
max_path_length=params['max_path_length'],
parallel=params['parallel'],
)
sample_processor = MetaSampleProcessor(
baseline=baseline,
discount=params['discount'],
gae_lambda=params['gae_lambda'],
normalize_adv=params['normalize_adv'],
)
algo = ProMP(
policy=policy,
inner_lr=params['inner_lr'],