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 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):
reward_baseline = LinearTimeBaseline()
return_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 = DiceMAMLSampleProcessor(
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
)
algo = VPG_DICEMAML(
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'], # This is repeated in MAMLPPO, it's confusing
)
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()
def run_experiment(**kwargs):
exp_dir = os.getcwd() + '/data/' + EXP_NAME + kwargs.get('exp_name', '')
logger.configure(dir=exp_dir,
format_strs=['csv', 'stdout', 'log'],
snapshot_mode='last')
json.dump(kwargs,
open(exp_dir + '/params.json', 'w'),
indent=2,
sort_keys=True,
cls=ClassEncoder)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = kwargs.get(
'gpu_frac', 0.95)
sess = tf.Session(config=config)
with sess.as_default() as sess:
# Instantiate classes
set_seed(kwargs['seed'])
baseline = kwargs['baseline']()
env = kwargs['env']() # Wrappers?
policy = MetaGaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
meta_batch_size=kwargs['meta_batch_size'],
hidden_sizes=kwargs['policy_hidden_sizes'],
learn_std=kwargs['policy_learn_std'],
hidden_nonlinearity=kwargs['policy_hidden_nonlinearity'],
output_nonlinearity=kwargs['policy_output_nonlinearity'],
)
dynamics_model = MLPDynamicsEnsemble(
'dynamics-ensemble',
env=env,
num_models=kwargs['num_models'],
hidden_nonlinearity=kwargs['dyanmics_hidden_nonlinearity'],
hidden_sizes=kwargs['dynamics_hidden_sizes'],
output_nonlinearity=kwargs['dyanmics_output_nonlinearity'],
learning_rate=kwargs['dynamics_learning_rate'],
batch_size=kwargs['dynamics_batch_size'],
buffer_size=kwargs['dynamics_buffer_size'],
rolling_average_persitency=kwargs['rolling_average_persitency'])
env_sampler = MetaSampler(
env=env,
policy=policy,
rollouts_per_meta_task=kwargs['real_env_rollouts_per_meta_task'],
meta_batch_size=kwargs['meta_batch_size'],
max_path_length=kwargs['max_path_length'],
parallel=kwargs['parallel'],
)
model_sampler = MBMPOSampler(
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'],
dynamics_model=dynamics_model,
deterministic=kwargs['deterministic'],
)
dynamics_sample_processor = ModelSampleProcessor(
baseline=baseline,
discount=kwargs['discount'],
gae_lambda=kwargs['gae_lambda'],
normalize_adv=kwargs['normalize_adv'],
positive_adv=kwargs['positive_adv'],
)
model_sample_processor = MAMLSampleProcessor(
baseline=baseline,
discount=kwargs['discount'],
gae_lambda=kwargs['gae_lambda'],
normalize_adv=kwargs['normalize_adv'],
positive_adv=kwargs['positive_adv'],
)
algo = TRPOMAML(
policy=policy,
step_size=kwargs['step_size'],
inner_type=kwargs['inner_type'],
inner_lr=kwargs['inner_lr'],
meta_batch_size=kwargs['meta_batch_size'],
num_inner_grad_steps=kwargs['num_inner_grad_steps'],
exploration=kwargs['exploration'],
)
trainer = Trainer(
algo=algo,
policy=policy,
env=env,
model_sampler=model_sampler,
env_sampler=env_sampler,
model_sample_processor=model_sample_processor,
dynamics_sample_processor=dynamics_sample_processor,
dynamics_model=dynamics_model,
num_rollouts_per_iter=int(kwargs['meta_batch_size'] *
kwargs['fraction_meta_batch_size']),
n_itr=kwargs['n_itr'],
num_inner_grad_steps=kwargs['num_inner_grad_steps'],
dynamics_model_max_epochs=kwargs['dynamics_max_epochs'],
log_real_performance=kwargs['log_real_performance'],
meta_steps_per_iter=kwargs['meta_steps_per_iter'],
sample_from_buffer=kwargs['sample_from_buffer'],
sess=sess,
)
trainer.train()
def run_base(exp_dir, **kwargs):
config = ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = kwargs.get(
'gpu_frac', 0.95)
# Instantiate classes
set_seed(kwargs['seed'])
baseline = kwargs['baseline']()
if kwargs['env'] == 'Ant':
env = AntEnv()
simulation_sleep = 0.05 * kwargs['num_rollouts'] * kwargs[
'max_path_length'] * kwargs['simulation_sleep_frac']
elif kwargs['env'] == 'HalfCheetah':
env = HalfCheetahEnv()
simulation_sleep = 0.05 * kwargs['num_rollouts'] * kwargs[
'max_path_length'] * kwargs['simulation_sleep_frac']
elif kwargs['env'] == 'Hopper':
env = HopperEnv()
simulation_sleep = 0.008 * kwargs['num_rollouts'] * kwargs[
'max_path_length'] * kwargs['simulation_sleep_frac']
elif kwargs['env'] == 'Walker2d':
env = Walker2dEnv()
simulation_sleep = 0.008 * kwargs['num_rollouts'] * kwargs[
'max_path_length'] * kwargs['simulation_sleep_frac']
else:
raise NotImplementedError
policy = MetaGaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
meta_batch_size=kwargs['meta_batch_size'],
hidden_sizes=kwargs['policy_hidden_sizes'],
learn_std=kwargs['policy_learn_std'],
hidden_nonlinearity=kwargs['policy_hidden_nonlinearity'],
output_nonlinearity=kwargs['policy_output_nonlinearity'],
)
dynamics_model = MLPDynamicsEnsemble(
'dynamics-ensemble',
env=env,
num_models=kwargs['num_models'],
hidden_nonlinearity=kwargs['dyanmics_hidden_nonlinearity'],
hidden_sizes=kwargs['dynamics_hidden_sizes'],
output_nonlinearity=kwargs['dyanmics_output_nonlinearity'],
learning_rate=kwargs['dynamics_learning_rate'],
batch_size=kwargs['dynamics_batch_size'],
buffer_size=kwargs['dynamics_buffer_size'],
rolling_average_persitency=kwargs['rolling_average_persitency'],
)
'''-------- dumps and reloads -----------------'''
baseline_pickle = pickle.dumps(baseline)
env_pickle = pickle.dumps(env)
receiver, sender = Pipe()
p = Process(
target=init_vars,
name="init_vars",
args=(sender, config, policy, dynamics_model),
daemon=True,
)
p.start()
policy_pickle, dynamics_model_pickle = receiver.recv()
receiver.close()
'''-------- following classes depend on baseline, env, policy, dynamics_model -----------'''
worker_data_feed_dict = {
'env_sampler': {
'rollouts_per_meta_task':
kwargs['real_env_rollouts_per_meta_task'],
'meta_batch_size': kwargs['meta_batch_size'],
'max_path_length': kwargs['max_path_length'],
'parallel': kwargs['parallel'],
},
'dynamics_sample_processor': {
'discount': kwargs['discount'],
'gae_lambda': kwargs['gae_lambda'],
'normalize_adv': kwargs['normalize_adv'],
'positive_adv': kwargs['positive_adv'],
},
}
worker_model_feed_dict = {}
worker_policy_feed_dict = {
'model_sampler': {
'rollouts_per_meta_task': kwargs['rollouts_per_meta_task'],
'meta_batch_size': kwargs['meta_batch_size'],
'max_path_length': kwargs['max_path_length'],
'deterministic': kwargs['deterministic'],
},
'model_sample_processor': {
'discount': kwargs['discount'],
'gae_lambda': kwargs['gae_lambda'],
'normalize_adv': kwargs['normalize_adv'],
'positive_adv': kwargs['positive_adv'],
},
'algo': {
'step_size': kwargs['step_size'],
'inner_type': kwargs['inner_type'],
'inner_lr': kwargs['inner_lr'],
'meta_batch_size': kwargs['meta_batch_size'],
'num_inner_grad_steps': kwargs['num_inner_grad_steps'],
'exploration': kwargs['exploration'],
}
}
trainer = ParallelTrainer(
exp_dir=exp_dir,
policy_pickle=policy_pickle,
env_pickle=env_pickle,
baseline_pickle=baseline_pickle,
dynamics_model_pickle=dynamics_model_pickle,
feed_dicts=[
worker_data_feed_dict, worker_model_feed_dict,
worker_policy_feed_dict
],
n_itr=kwargs['n_itr'],
num_inner_grad_steps=kwargs['num_inner_grad_steps'],
initial_random_samples=kwargs['initial_random_samples'],
flags_need_query=kwargs['flags_need_query'],
num_rollouts_per_iter=int(kwargs['meta_batch_size'] *
kwargs['fraction_meta_batch_size']),
config=config,
simulation_sleep=simulation_sleep,
)
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)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = kwargs.get(
'gpu_frac', 0.95)
sess = tf.Session(config=config)
with sess.as_default() as sess:
# Instantiate classes
set_seed(kwargs['seed'])
baseline = kwargs['baseline']()
env = normalize(kwargs['env']()) # Wrappers?
policy = MetaGaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
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 = MetaSampler(
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'],
)
sample_processor = MAMLSampleProcessor(
baseline=baseline,
discount=kwargs['discount'],
gae_lambda=kwargs['gae_lambda'],
normalize_adv=kwargs['normalize_adv'],
positive_adv=kwargs['positive_adv'],
)
algo = TRPOMAML(
policy=policy,
step_size=kwargs['step_size'],
inner_type=kwargs['inner_type'],
inner_lr=kwargs['inner_lr'],
meta_batch_size=kwargs['meta_batch_size'],
num_inner_grad_steps=kwargs['num_inner_grad_steps'],
exploration=kwargs['exploration'],
)
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'],
sess=sess,
)
trainer.train()
def run_experiment(**config):
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(config,
open(exp_dir + '/params.json', 'w'),
indent=2,
sort_keys=True,
cls=ClassEncoder)
# Instantiate classes
set_seed(config['seed'])
baseline = config['baseline']()
env = normalize(config['env']()) # Wrappers?
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'],
learn_std=config['learn_std'],
hidden_nonlinearity=config['hidden_nonlinearity'],
output_nonlinearity=config['output_nonlinearity'],
)
# Load policy here
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 = 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'],
)
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'])
baseline = kwargs['baseline']()
env = normalize(kwargs['env']()) # Wrappers?
policy = MetaGaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
meta_batch_size=kwargs['meta_batch_size'],
hidden_sizes=kwargs['policy_hidden_sizes'],
learn_std=kwargs['policy_learn_std'],
hidden_nonlinearity=kwargs['policy_hidden_nonlinearity'],
output_nonlinearity=kwargs['policy_output_nonlinearity'],
)
dynamics_model = MLPDynamicsEnsemble('dynamics-ensemble',
env=env,
num_models=kwargs['num_models'],
hidden_nonlinearity=kwargs['dyanmics_hidden_nonlinearity'],
hidden_sizes=kwargs['dynamics_hidden_sizes'],
output_nonlinearity=kwargs['dyanmics_output_nonlinearity'],
learning_rate=kwargs['dynamics_learning_rate'],
batch_size=kwargs['dynamics_batch_size'],
buffer_size=kwargs['dynamics_buffer_size'],
)
env_sampler = SingleMetaSampler(
env=env,
policy=policy,
rollouts_per_meta_task=kwargs['real_env_rollouts_per_meta_task'],
meta_batch_size=kwargs['meta_batch_size'],
max_path_length=kwargs['max_path_length'],
parallel=kwargs['parallel'],
)
model_sampler = MBMPOSampler(
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'],
dynamics_model=dynamics_model,
)
dynamics_sample_processor = ModelSampleProcessor(
baseline=baseline,
discount=kwargs['discount'],
gae_lambda=kwargs['gae_lambda'],
normalize_adv=kwargs['normalize_adv'],
positive_adv=kwargs['positive_adv'],
)
model_sample_processor = MAMLSampleProcessor(
baseline=baseline,
discount=kwargs['discount'],
gae_lambda=kwargs['gae_lambda'],
normalize_adv=kwargs['normalize_adv'],
positive_adv=kwargs['positive_adv'],
)
algo = TRPOMAML(
policy=policy,
step_size=kwargs['step_size'],
inner_type=kwargs['inner_type'],
inner_lr=kwargs['inner_lr'],
meta_batch_size=kwargs['meta_batch_size'],
num_inner_grad_steps=kwargs['num_inner_grad_steps'],
exploration=kwargs['exploration'],
)
trainer = Trainer(
algo=algo,
policy=policy,
env=env,
model_sampler=model_sampler,
env_sampler=env_sampler,
model_sample_processor=model_sample_processor,
dynamics_sample_processor=dynamics_sample_processor,
dynamics_model=dynamics_model,
n_itr=kwargs['n_itr'],
num_inner_grad_steps=kwargs['num_inner_grad_steps'],
dynamics_model_max_epochs=kwargs['dynamics_max_epochs'],
log_real_performance=kwargs['log_real_performance'],
meta_steps_per_iter=kwargs['meta_steps_per_iter'],
initial_random_samples=True,
sample_from_buffer=True,
)
trainer.train()
