def set_shared_params(self, state):
MLPDynamicsEnsemble.set_shared_params(self, state)
sess = tf.get_default_session()
sess.run(self._assign_ops_var,
feed_dict={
self._min_log_var_ph: state['min_log_var'],
self._max_log_var_ph: state['max_log_var'],
})
def __setstate__(self, state):
MLPDynamicsEnsemble.__setstate__(self, state)
sess = tf.get_default_session()
sess.run(self._assign_ops_var,
feed_dict={
self._min_log_var_ph: state['min_log_var'],
self._max_log_var_ph: state['max_log_var']
})
def run_experiment(**kwargs):
exp_dir = os.getcwd() + '/data/parallel_mb_ppo/' + EXP_NAME + '/' + kwargs.get('exp_name', '')
logger.configure(dir=exp_dir, format_strs=['stdout', 'log', 'csv'], 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 = normalize(kwargs['env']()) # Wrappers?
policy = GaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
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 = Sampler(
env=env,
policy=policy,
num_rollouts=kwargs['num_rollouts'],
max_path_length=kwargs['max_path_length'],
n_parallel=kwargs['n_parallel'],
)
model_sampler = METRPOSampler(
env=env,
policy=policy,
num_rollouts=kwargs['imagined_num_rollouts'],
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 = SampleProcessor(
baseline=baseline,
discount=kwargs['discount'],
gae_lambda=kwargs['gae_lambda'],
normalize_adv=kwargs['normalize_adv'],
positive_adv=kwargs['positive_adv'],
)
algo = PPO(
policy=policy,
learning_rate=kwargs['learning_rate'],
clip_eps=kwargs['clip_eps'],
max_epochs=kwargs['num_ppo_steps'],
)
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'],
dynamics_model_max_epochs=kwargs['dynamics_max_epochs'],
log_real_performance=kwargs['log_real_performance'],
steps_per_iter=kwargs['steps_per_iter'],
sample_from_buffer=True,
sess=sess,
)
trainer.train()
def run_experiment(**kwargs):
print()
exp_dir = os.getcwd(
) + '/data/parallel_mb_ppo/' + EXP_NAME + '/' + kwargs.get('exp_name', '')
print("\n---------- experiment with dir {} ---------------------------".
format(exp_dir))
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 = 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']()
env = normalize(kwargs['env']()) # Wrappers?
policy = GaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
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'],
)
'''-------- 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=False,
)
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': {
'num_rollouts': kwargs['num_rollouts'],
'max_path_length': kwargs['max_path_length'],
'n_parallel': kwargs['n_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': {
'num_rollouts': kwargs['imagined_num_rollouts'],
'max_path_length': kwargs['max_path_length'],
'dynamics_model': dynamics_model,
'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': {
'learning_rate': kwargs['learning_rate'],
'clip_eps': kwargs['clip_eps'],
'max_epochs': kwargs['num_ppo_steps'],
}
}
trainer = ParallelTrainer(
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'],
dynamics_model_max_epochs=kwargs['dynamics_max_epochs'],
log_real_performance=kwargs['log_real_performance'],
steps_per_iter=kwargs['steps_per_iter'],
flags_need_query=kwargs['flags_need_query'],
config=config,
simulation_sleep=kwargs['simulation_sleep'],
)
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 = normalize(AntEnv())
simulation_sleep = 0.05 * kwargs['num_rollouts'] * kwargs['max_path_length'] * kwargs['simulation_sleep_frac']
elif kwargs['env'] == 'HalfCheetah':
env = normalize(HalfCheetahEnv())
simulation_sleep = 0.05 * kwargs['num_rollouts'] * kwargs['max_path_length'] * kwargs['simulation_sleep_frac']
elif kwargs['env'] == 'Hopper':
env = normalize(HopperEnv())
simulation_sleep = 0.008 * kwargs['num_rollouts'] * kwargs['max_path_length'] * kwargs['simulation_sleep_frac']
elif kwargs['env'] == 'Walker2d':
env = normalize(Walker2dEnv())
simulation_sleep = 0.008 * kwargs['num_rollouts'] * kwargs['max_path_length'] * kwargs['simulation_sleep_frac']
else:
raise NotImplementedError
policy = GaussianMLPPolicy(
name="meta-policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
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': {
'num_rollouts': kwargs['num_rollouts'],
'max_path_length': kwargs['max_path_length'],
'n_parallel': kwargs['n_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': {
'num_rollouts': kwargs['imagined_num_rollouts'],
'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': {
'learning_rate': kwargs['learning_rate'],
'clip_eps': kwargs['clip_eps'],
'max_epochs': kwargs['num_ppo_steps'],
}
}
trainer = ParallelTrainer(
exp_dir=exp_dir,
algo_str=kwargs['algo'],
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'],
flags_need_query=kwargs['flags_need_query'],
config=config,
simulation_sleep=simulation_sleep,
sampler_str=kwargs['sampler'],
)
trainer.train()
def run_base(exp_dir, **config):
config_sess = ConfigProto()
config_sess.gpu_options.allow_growth = True
config_sess.gpu_options.per_process_gpu_memory_fraction = config.get('gpu_frac', 0.95)
# Instantiate classes
set_seed(config['seed'])
if config['env'] == 'Ant':
env = AntEnv()
simulation_sleep = 0.05 * config['num_rollouts'] * config['max_path_length'] * config['simulation_sleep_frac']
elif config['env'] == 'HalfCheetah':
env = HalfCheetahEnv()
simulation_sleep = 0.05 * config['num_rollouts'] * config['max_path_length'] * config['simulation_sleep_frac']
elif config['env'] == 'Hopper':
env = HopperEnv()
simulation_sleep = 0.008 * config['num_rollouts'] * config['max_path_length'] * config['simulation_sleep_frac']
elif config['env'] == 'Walker2d':
env = Walker2dEnv()
simulation_sleep = 0.008 * config['num_rollouts'] * config['max_path_length'] * config['simulation_sleep_frac']
else:
raise NotImplementedError
if config['probabilistic_dynamics']:
dynamics_model = ProbMLPDynamicsEnsemble(
'prob-dynamics-ensemble',
env=env,
num_models=config['num_models'],
hidden_nonlinearity=config['dynamics_hidden_nonlinearity'],
hidden_sizes=config['dynamics_hidden_sizes'],
output_nonlinearity=config['dynamics_output_nonlinearity'],
learning_rate=config['dynamics_learning_rate'],
batch_size=config['dynamics_batch_size'],
buffer_size=config['dynamics_buffer_size'],
rolling_average_persitency=config['rolling_average_persitency']
)
else:
dynamics_model = MLPDynamicsEnsemble(
'dynamics-ensemble',
env=env,
num_models=config['num_models'],
hidden_nonlinearity=config['dynamics_hidden_nonlinearity'],
hidden_sizes=config['dynamics_hidden_sizes'],
output_nonlinearity=config['dynamics_output_nonlinearity'],
learning_rate=config['dynamics_learning_rate'],
batch_size=config['dynamics_batch_size'],
buffer_size=config['dynamics_buffer_size'],
rolling_average_persitency=config['rolling_average_persitency']
)
policy = MPCController(
name="policy",
env=env,
dynamics_model=dynamics_model,
discount=config['discount'],
n_candidates=config['n_candidates'],
horizon=config['horizon'],
use_cem=config['use_cem'],
num_cem_iters=config['num_cem_iters'],
)
'''-------- dumps and reloads -----------------'''
env_pickle = pickle.dumps(env)
receiver, sender = Pipe()
p = Process(
target=init_vars,
name="init_vars",
args=(sender, config_sess, policy, dynamics_model),
daemon=False,
)
p.start()
policy_pickle, dynamics_model_pickle = receiver.recv()
receiver.close()
'''-------- following classes depend on baseline, env, policy, dynamics_model -----------'''
worker_data_feed_dict = {
'sampler': {
'num_rollouts': config['num_rollouts'],
'max_path_length': config['max_path_length'],
'n_parallel': config['n_parallel'],
},
'sample_processor': {},
}
worker_model_feed_dict = {}
trainer = ParallelTrainer(
exp_dir=exp_dir,
env_pickle=env_pickle,
policy_pickle=policy_pickle,
baseline_pickle=None,
dynamics_model_pickle=dynamics_model_pickle,
feed_dicts=[worker_data_feed_dict, worker_model_feed_dict],
n_itr=config['n_itr'],
initial_random_samples=config['initial_random_samples'],
initial_sinusoid_samples=config['initial_sinusoid_samples'],
flags_need_query=config['flags_need_query'],
config=config_sess,
simulation_sleep=simulation_sleep,
)
trainer.train()
def run_experiment(**config):
exp_dir = os.getcwd() + '/data/' + EXP_NAME + '/' + config.get('exp_name', '')
logger.configure(dir=exp_dir, format_strs=['stdout', 'log', 'csv'], snapshot_mode='last')
json.dump(config, open(exp_dir + '/params.json', 'w'), indent=2, sort_keys=True, cls=ClassEncoder)
config_sess = tf.ConfigProto()
config_sess.gpu_options.allow_growth = True
config_sess.gpu_options.per_process_gpu_memory_fraction = config.get('gpu_frac', 0.95)
sess = tf.Session(config=config_sess)
with sess.as_default() as sess:
env = config['env']()
if config['recurrent']:
dynamics_model = RNNDynamicsEnsemble(
name="dyn_model",
env=env,
hidden_sizes=config['hidden_sizes_model'],
learning_rate=config['learning_rate'],
backprop_steps=config['backprop_steps'],
cell_type=config['cell_type'],
num_models=config['num_models'],
batch_size=config['batch_size_model'],
normalize_input=True,
)
policy = RNNMPCController(
name="policy",
env=env,
dynamics_model=dynamics_model,
discount=config['discount'],
n_candidates=config['n_candidates'],
horizon=config['horizon'],
use_cem=config['use_cem'],
num_cem_iters=config['num_cem_iters'],
use_reward_model=config['use_reward_model']
)
else:
dynamics_model = MLPDynamicsEnsemble(
name="dyn_model",
env=env,
learning_rate=config['learning_rate'],
hidden_sizes=config['hidden_sizes_model'],
weight_normalization=config['weight_normalization_model'],
num_models=config['num_models'],
valid_split_ratio=config['valid_split_ratio'],
rolling_average_persitency=config['rolling_average_persitency'],
hidden_nonlinearity=config['hidden_nonlinearity_model'],
batch_size=config['batch_size_model'],
)
policy = MPCController(
name="policy",
env=env,
dynamics_model=dynamics_model,
discount=config['discount'],
n_candidates=config['n_candidates'],
horizon=config['horizon'],
use_cem=config['use_cem'],
num_cem_iters=config['num_cem_iters'],
)
sampler = Sampler(
env=env,
policy=policy,
num_rollouts=config['num_rollouts'],
max_path_length=config['max_path_length'],
n_parallel=config['n_parallel'],
)
sample_processor = ModelSampleProcessor()
algo = Trainer(
env=env,
policy=policy,
dynamics_model=dynamics_model,
sampler=sampler,
dynamics_sample_processor=sample_processor,
n_itr=config['n_itr'],
initial_random_samples=config['initial_random_samples'],
dynamics_model_max_epochs=config['dynamic_model_epochs'],
initial_sinusoid_samples=config['initial_sinusoid_samples'],
sess=sess,
)
algo.train()
def run_experiment(**kwargs):
exp_dir = os.getcwd() + '/data/' + EXP_NAME + '/' + kwargs.get(
'exp_name', '')
logger.configure(dir=exp_dir,
format_strs=['stdout', 'log', 'csv'],
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']()
if not kwargs['use_images']:
env = normalize(kwargs['env']())
else:
vae = VAE(latent_dim=8)
env = image_wrapper(normalize(kwargs['env']()),
vae=vae,
latent_dim=32)
policy = NNPolicy(
name="policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
hidden_sizes=kwargs['hidden_sizes'],
normalization=None,
)
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'],
)
# dynamics_model = None
assert kwargs['rollouts_per_policy'] % kwargs['num_models'] == 0
env_sampler = Sampler(
env=env,
policy=policy,
num_rollouts=kwargs['num_rollouts'],
max_path_length=kwargs['max_path_length'],
n_parallel=kwargs['num_rollouts'],
)
# TODO: I'm not sure if it works with more than one rollout per model
model_sampler = ARSSampler(
env=env,
policy=policy,
dynamics_model=dynamics_model,
rollouts_per_policy=kwargs['rollouts_per_policy'],
max_path_length=kwargs['horizon'],
num_deltas=kwargs['num_deltas'],
n_parallel=1,
)
dynamics_sample_processor = ModelSampleProcessor(
baseline=baseline,
discount=kwargs['discount'],
gae_lambda=kwargs['gae_lambda'],
normalize_adv=kwargs['normalize_adv'],
positive_adv=kwargs['positive_adv'],
)
ars_sample_processor = ARSSamplerProcessor(
baseline=baseline,
discount=kwargs['discount'],
gae_lambda=kwargs['gae_lambda'],
normalize_adv=kwargs['normalize_adv'],
positive_adv=kwargs['positive_adv'],
uncertainty_coeff=kwargs['uncertainty_coeff'])
algo = RandomSearchOptimizer(policy=policy,
learning_rate=kwargs['learning_rate'],
num_deltas=kwargs['num_deltas'],
percentile=kwargs['percentile'])
trainer = Trainer(
algo=algo,
policy=policy,
env=env,
model_sampler=model_sampler,
env_sampler=env_sampler,
ars_sample_processor=ars_sample_processor,
dynamics_sample_processor=dynamics_sample_processor,
dynamics_model=dynamics_model,
num_deltas=kwargs['num_deltas'],
n_itr=kwargs['n_itr'],
dynamics_model_max_epochs=kwargs['dynamics_max_epochs'],
log_real_performance=kwargs['log_real_performance'],
steps_per_iter=kwargs['steps_per_iter'],
delta_std=kwargs['delta_std'],
sess=sess,
initial_random_samples=True,
sample_from_buffer=kwargs['sample_from_buffer'])
trainer.train()
def get_shared_param_values(self): # to feed policy
state = MLPDynamicsEnsemble.get_shared_param_values(self)
sess = tf.get_default_session()
state['min_log_var'] = sess.run(self.min_logvar)
state['max_log_var'] = sess.run(self.max_logvar)
return state
def __getstate__(self):
state = MLPDynamicsEnsemble.__getstate__(self)
sess = tf.get_default_session()
state['min_log_var'] = sess.run(self.min_logvar)
state['max_log_var'] = sess.run(self.max_logvar)
return state
def run_experiment(**kwargs):
num = Num()
exp_name = EXP_NAME + str(num.EXP_NUM)
exp_dir = os.getcwd() + '/data/video_peg/' + EXP_NAME + kwargs.get(
'exp_name', '')
logger.configure(dir=exp_dir,
format_strs=['csv', 'stdout', 'log'],
snapshot_mode='all') #change to all
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)
Num.EXP_NUM += 1
with sess.as_default() as sess:
# Instantiate classesLogger
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 = BaseSampler(
env=env,
policy=policy,
# rollouts_per_meta_task=kwargs['real_env_rollouts_per_meta_task'],
num_rollouts=kwargs['meta_batch_size'],
max_path_length=kwargs['max_path_length'],
sleep_reset=2.5,
#parallel=kwargs['parallel'],
# parallel=False
)
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,
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=True,
sess=sess,
)
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()
