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 = MAMLSampler(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 = MAMLSampler(self.random_env,
self.random_policy,
self.batch_size,
self.meta_batch_size,
self.path_length,
parallel=True)
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 = MAMLSampler(
env=env,
policy=policy,
rollouts_per_meta_task=2,
meta_batch_size=10,
max_path_length=50,
parallel=False,
)
self.sample_processor = MAMLSampleProcessor(
baseline=baseline,
discount=0.99,
gae_lambda=1.0,
normalize_adv=True,
positive_adv=False,
)
self.algo = PPOMAML(
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,
clip_outer=True,
target_outer_step=0,
target_inner_step=2e-2,
init_outer_kl_penalty=0,
init_inner_kl_penalty=1e-3,
adaptive_outer_kl_penalty=False,
adaptive_inner_kl_penalty=True,
anneal_factor=1.0,
)
def main(config):
baseline = LinearFeatureBaseline()
# env = normalize(HalfCheetahRandDirecEnv())
env = HopperRandParamsEnv(3.5)
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 = MAMLSampler(
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 = MAMLSampleProcessor(
baseline=baseline,
discount=config['discount'],
gae_lambda=config['gae_lambda'],
normalize_adv=config['normalize_adv'],
positive_adv=config['positive_adv'],
)
algo = PPOMAML(
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_ppo_steps'],
num_minibatches=config['num_minibatches'],
clip_eps=config['clip_eps'],
clip_outer=config['clip_outer'],
target_outer_step=config['target_outer_step'],
target_inner_step=config['target_inner_step'],
init_outer_kl_penalty=config['init_outer_kl_penalty'],
init_inner_kl_penalty=config['init_inner_kl_penalty'],
adaptive_outer_kl_penalty=config['adaptive_outer_kl_penalty'],
adaptive_inner_kl_penalty=config['adaptive_inner_kl_penalty'],
anneal_factor=config['anneal_factor'],
)
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'], # This is repeated in MAMLPPO, it's confusing
)
trainer.train()
def main(config):
baseline = LinearFeatureBaseline()
env = normalize(HalfCheetahRandDirecEnv())
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 = MAMLSampler(
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 = MAMLSampleProcessor(
baseline=baseline,
discount=config['discount'],
gae_lambda=config['gae_lambda'],
normalize_adv=config['normalize_adv'],
positive_adv=config['positive_adv'],
)
algo = TRPOMAML(policy=policy,
step_size=config['step_size'],
inner_type=config['inner_type'],
meta_batch_size=config['meta_batch_size'],
num_inner_grad_steps=config['num_inner_grad_steps'],
inner_lr=config['inner_lr'])
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'], # This is repeated in MAMLPPO, it's confusing
)
trainer.train()
def main(config):
baseline = LinearFeatureBaseline()
env = normalize(HopperRandParamsEnv())
obs_dim = np.prod(env.observation_space.shape)
policy = GaussianMLPPolicy(
name="meta-policy",
obs_dim=obs_dim,
action_dim=np.prod(env.action_space.shape),
meta_batch_size=config['meta_batch_size'],
hidden_sizes=config['hidden_sizes'],
)
sampler = MAMLSampler(
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'],
envs_per_task=5,
)
sample_processor = SingleSampleProcessor(
baseline=baseline,
discount=config['discount'],
gae_lambda=config['gae_lambda'],
normalize_adv=config['normalize_adv'],
positive_adv=config['positive_adv'],
)
algo = PPO(policy=policy,
learning_rate=config['learning_rate'],
max_epochs=config['max_epochs'])
trainer = Trainer(
algo=algo,
policy=policy,
env=env,
sampler=sampler,
sample_processor=sample_processor,
n_itr=config['n_itr'],
)
trainer.train()
def run_experiment(**kwargs):
exp_dir = os.getcwd() + '/data/' + EXP_NAME
logger.configure(dir=exp_dir,
format_strs=['stdout', 'log', 'csv'],
snapshot_mode='last_gap',
snapshot_gap=50)
json.dump(kwargs,
open(exp_dir + '/params.json', 'w'),
indent=2,
sort_keys=True,
cls=ClassEncoder)
# Instantiate classes
set_seed(kwargs['seed'])
reward_baseline = LinearTimeBaseline()
return_baseline = LinearFeatureBaseline()
env = normalize(kwargs['env']()) # Wrappers?
policy = MetaGaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape), # Todo...?
action_dim=np.prod(env.action_space.shape),
meta_batch_size=kwargs['meta_batch_size'],
hidden_sizes=kwargs['hidden_sizes'],
learn_std=kwargs['learn_std'],
hidden_nonlinearity=kwargs['hidden_nonlinearity'],
output_nonlinearity=kwargs['output_nonlinearity'],
)
# Load policy here
sampler = MAMLSampler(
env=env,
policy=policy,
rollouts_per_meta_task=kwargs['rollouts_per_meta_task'],
meta_batch_size=kwargs['meta_batch_size'],
max_path_length=kwargs['max_path_length'],
parallel=kwargs['parallel'],
envs_per_task=int(kwargs['rollouts_per_meta_task'] / 2))
sample_processor = DiceMAMLSampleProcessor(
baseline=reward_baseline,
max_path_length=kwargs['max_path_length'],
discount=kwargs['discount'],
normalize_adv=kwargs['normalize_adv'],
positive_adv=kwargs['positive_adv'],
return_baseline=return_baseline)
algo = VPG_DICEMAML(policy=policy,
max_path_length=kwargs['max_path_length'],
meta_batch_size=kwargs['meta_batch_size'],
num_inner_grad_steps=kwargs['num_inner_grad_steps'],
inner_lr=kwargs['inner_lr'],
learning_rate=kwargs['learning_rate'])
trainer = Trainer(
algo=algo,
policy=policy,
env=env,
sampler=sampler,
sample_processor=sample_processor,
n_itr=kwargs['n_itr'],
num_inner_grad_steps=kwargs['num_inner_grad_steps'],
)
trainer.train()
args = parser.parse_args(sys.argv[1:])
sess = tf.InteractiveSession()
policy = joblib.load(args.policy)['policy']
policy.switch_to_pre_update()
baseline = LinearFeatureBaseline()
env = normalize(AntRandGoalEnv())
sampler = MAMLSampler(
env=env,
policy=policy,
rollouts_per_meta_task=BATCH_SIZE,
meta_batch_size=META_BATCH_SIZE,
max_path_length=PATH_LENGTH,
parallel=True,
envs_per_task=20,
)
sample_processor = MAMLSampleProcessor(
baseline=baseline,
discount=0.99,
gae_lambda=1,
normalize_adv=True,
positive_adv=False,
)
# Doesn't matter which algo
algo = VPGMAML(
def run_experiment(**kwargs):
exp_dir = os.getcwd() + '/data/' + EXP_NAME
logger.configure(dir=exp_dir,
format_strs=['stdout', 'log', 'csv'],
snapshot_mode='last_gap',
snapshot_gap=50)
json.dump(kwargs,
open(exp_dir + '/params.json', 'w'),
indent=2,
sort_keys=True,
cls=ClassEncoder)
# Instantiate classes
set_seed(kwargs['seed'])
sess = tf.Session()
with sess.as_default() as sess:
config = json.load(open(osp.join(kwargs['path'], 'params.json'), 'r'))
data = joblib.load(osp.join(kwargs['path'], 'params.pkl'))
policy = data['policy']
env = data['env']
baseline = data['baseline']
if kwargs['rollouts_per_meta_task'] is None:
rollouts_per_meta_task = int(
np.ceil(config['rollouts_per_meta_task'] /
config['meta_batch_size']))
else:
rollouts_per_meta_task = kwargs['rollouts_per_meta_task']
sampler = MAMLSampler(
env=env,
policy=policy,
rollouts_per_meta_task=rollouts_per_meta_task,
meta_batch_size=config['meta_batch_size'],
max_path_length=kwargs['max_path_length'],
parallel=kwargs['parallel'],
)
sample_processor = SampleProcessor(
baseline=baseline,
discount=config['discount'],
normalize_adv=config['normalize_adv'],
positive_adv=config['positive_adv'],
)
algo = VPG(
policy=policy,
learning_rate=config['inner_lr'],
)
tester = Tester(
algo=algo,
policy=policy,
env=env,
sampler=sampler,
sample_processor=sample_processor,
n_itr=kwargs['n_itr'],
sess=sess,
task=None,
)
tester.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 = MAMLSampler(
env=env,
policy=policy,
rollouts_per_meta_task=2,
meta_batch_size=10,
max_path_length=50,
parallel=False,
)
self.sample_processor = MAMLSampleProcessor(
baseline=baseline,
discount=0.99,
gae_lambda=1.0,
normalize_adv=True,
positive_adv=False,
)
self.algo = PPOMAML(
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,
clip_outer=True,
target_outer_step=0,
target_inner_step=2e-2,
init_outer_kl_penalty=0,
init_inner_kl_penalty=1e-3,
adaptive_outer_kl_penalty=False,
adaptive_inner_kl_penalty=True,
anneal_factor=1.0,
)
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))