def prepare_start(self, policy_pickle, env_pickle, baseline_pickle,
feed_dict, config, initial_random_samples):
import tensorflow as tf
self.sess = sess = tf.Session(config=config)
with sess.as_default():
""" --------------------- Construct instances -------------------"""
from asynch_mb.samplers.sampler import Sampler
from asynch_mb.samplers.mb_sample_processor import ModelSampleProcessor
env = pickle.loads(env_pickle)
policy = pickle.loads(policy_pickle)
baseline = pickle.loads(baseline_pickle)
sess.run(tf.initializers.global_variables())
self.env = env
self.env_sampler = Sampler(env=env,
policy=policy,
**feed_dict['env_sampler'])
self.dynamics_sample_processor = ModelSampleProcessor(
baseline=baseline, **feed_dict['dynamics_sample_processor'])
""" ------------------- Step and Push ------------------"""
samples_data = self.step(random=initial_random_samples)
self.push(samples_data)
return 1
def construct_from_feed_dict(
self,
policy_pickle,
env_pickle,
baseline_pickle, # UNUSED
dynamics_model_pickle,
feed_dict,
):
from asynch_mb.samplers.sampler import Sampler
from asynch_mb.samplers.mb_sample_processor import ModelSampleProcessor
env = pickle.loads(env_pickle)
policy = pickle.loads(policy_pickle)
self.env = env
self.env_sampler = Sampler(env=env,
policy=policy,
**feed_dict['sampler'])
self.dynamics_sample_processor = ModelSampleProcessor(
**feed_dict['sample_processor'])
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']()
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'],
rolling_average_persitency=kwargs['rolling_average_persitency']
)
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,
dynamics_model=dynamics_model,
num_rollouts=kwargs['imagined_num_rollouts'],
max_path_length=kwargs['max_path_length'],
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 = TRPO(
policy=policy,
step_size=kwargs['step_size'],
)
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=kwargs['sample_from_buffer'],
sess=sess,
)
trainer.train()
class WorkerData(Worker):
def __init__(self, simulation_sleep, video=False):
if video:
super().__init__(snapshot_mode='gap', snapshot_gap=int(30/1250/simulation_sleep)) # FIXME
else:
super().__init__()
self.simulation_sleep = simulation_sleep
self.env = None
self.env_sampler = None
self.dynamics_sample_processor = None
self.samples_data_arr = []
def construct_from_feed_dict(
self,
policy_pickle,
env_pickle,
baseline_pickle,
dynamics_model_pickle,
feed_dict
):
from asynch_mb.samplers.sampler import Sampler
from asynch_mb.samplers.mb_sample_processor import ModelSampleProcessor
env = pickle.loads(env_pickle)
policy = pickle.loads(policy_pickle)
baseline = pickle.loads(baseline_pickle)
self.env = env
self.env_sampler = Sampler(env=env, policy=policy, **feed_dict['env_sampler'])
self.dynamics_sample_processor = ModelSampleProcessor(
baseline=baseline,
**feed_dict['dynamics_sample_processor']
)
def prepare_start(self):
initial_random_samples = self.queue.get()
self.step(initial_random_samples)
self.push()
def step(self, random=False):
time_step = time.time()
'''------------- Obtaining samples from the environment -----------'''
if self.verbose:
logger.log("Data is obtaining samples...")
env_paths = self.env_sampler.obtain_samples(
log=True,
random=random,
log_prefix='Data-EnvSampler-',
)
'''-------------- Processing environment samples -------------------'''
if self.verbose:
logger.log("Data is processing environment samples...")
samples_data = self.dynamics_sample_processor.process_samples(
env_paths,
log=True,
log_prefix='Data-EnvTrajs-',
)
self.samples_data_arr.append(samples_data)
time_step = time.time() - time_step
time_sleep = max(self.simulation_sleep - time_step, 0)
time.sleep(time_sleep)
logger.logkv('Data-TimeStep', time_step)
logger.logkv('Data-TimeSleep', time_sleep)
# save snapshot
params = self.get_itr_snapshot()
logger.save_itr_params(self.itr_counter, params)
def _synch(self, policy_state_pickle):
time_synch = time.time()
policy_state = pickle.loads(policy_state_pickle)
assert isinstance(policy_state, dict)
self.env_sampler.policy.set_shared_params(policy_state)
time_synch = time.time() - time_synch
logger.logkv('Data-TimeSynch', time_synch)
def push(self):
time_push = time.time()
self.queue_next.put(pickle.dumps(self.samples_data_arr))
self.samples_data_arr = []
time_push = time.time() - time_push
logger.logkv('Data-TimePush', time_push)
def set_stop_cond(self):
if self.itr_counter >= self.n_itr:
self.stop_cond.set()
def get_itr_snapshot(self):
"""
Gets the current policy and env for storage
"""
return dict(itr=self.itr_counter, policy=self.env_sampler.policy, env=self.env)
def run_experiment(**kwargs):
exp_dir = os.getcwd() + '/data/' + 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']())
policy = GaussianMLPPolicy(
name="policy",
obs_dim=np.prod(env.observation_space.shape),
action_dim=np.prod(env.action_space.shape),
hidden_sizes=kwargs['hidden_sizes'],
learn_std=kwargs['learn_std'],
hidden_nonlinearity=kwargs['hidden_nonlinearity'],
output_nonlinearity=kwargs['output_nonlinearity'],
init_std=kwargs['init_std'],
squashed=kwargs['squashed'])
# Load policy here
sampler = Sampler(
env=env,
policy=policy,
num_rollouts=kwargs['num_rollouts'],
max_path_length=kwargs['max_path_length'],
n_parallel=kwargs['n_parallel'],
)
sample_processor = SingleSampleProcessor(
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'],
entropy_bonus=kwargs['entropy_bonus'],
)
trainer = Trainer(
algo=algo,
policy=policy,
env=env,
sampler=sampler,
sample_processor=sample_processor,
n_itr=kwargs['n_itr'],
sess=sess,
)
trainer.train()
class WorkerData(Worker):
def __init__(self, policy_ps, data_buffers, time_sleep, name, exp_dir,
n_itr, stop_cond):
super().__init__(name, exp_dir, n_itr, stop_cond)
self.policy_ps = policy_ps
self.data_buffers = data_buffers
self.time_sleep = time_sleep
self.env = None
self.env_sampler = None
self.dynamics_sample_processor = None
def prepare_start(self, policy_pickle, env_pickle, baseline_pickle,
feed_dict, config, initial_random_samples):
import tensorflow as tf
self.sess = sess = tf.Session(config=config)
with sess.as_default():
""" --------------------- Construct instances -------------------"""
from asynch_mb.samplers.sampler import Sampler
from asynch_mb.samplers.mb_sample_processor import ModelSampleProcessor
env = pickle.loads(env_pickle)
policy = pickle.loads(policy_pickle)
baseline = pickle.loads(baseline_pickle)
sess.run(tf.initializers.global_variables())
self.env = env
self.env_sampler = Sampler(env=env,
policy=policy,
**feed_dict['env_sampler'])
self.dynamics_sample_processor = ModelSampleProcessor(
baseline=baseline, **feed_dict['dynamics_sample_processor'])
""" ------------------- Step and Push ------------------"""
samples_data = self.step(random=initial_random_samples)
self.push(samples_data)
return 1
def step_wrapper(self):
self.pull()
samples_data = self.step()
self.push(samples_data)
return 1, 1
def step(self, random=False):
time_step = time.time()
'''------------- Obtaining samples from the environment -----------'''
if self.verbose:
logger.log("Data is obtaining samples...")
env_paths = self.env_sampler.obtain_samples(
log=True,
random=random,
log_prefix='Data-EnvSampler-',
)
'''-------------- Processing environment samples -------------------'''
if self.verbose:
logger.log("Data is processing environment samples...")
samples_data = self.dynamics_sample_processor.process_samples(
env_paths,
log=True,
log_prefix='Data-EnvTrajs-',
)
time_step = time.time() - time_step
time_sleep = max(self.time_sleep - time_step, 0)
time.sleep(time_sleep)
logger.logkv('Data-TimeStep', time_step)
logger.logkv('Data-TimeSleep', time_sleep)
return samples_data
def pull(self):
time_synch = time.time()
policy_params = ray.get(self.policy_ps.pull.remote())
assert isinstance(policy_params, dict)
self.env_sampler.policy.set_shared_params(policy_params)
logger.logkv('Data-TimePull', time.time() - time_synch)
def push(self, samples_data):
time_push = time.time()
# broadcast samples to all data buffers
samples_data_id = ray.put(samples_data)
for data_buffer in self.data_buffers:
# ray.get(data_buffer.push.remote(samples_data))
data_buffer.push.remote(samples_data_id)
logger.logkv('Data-TimePush', time.time() - time_push)
def set_stop_cond(self):
if self.step_counter >= self.n_itr:
ray.get(self.stop_cond.set.remote())
class WorkerData(Worker):
def __init__(self, simulation_sleep):
super().__init__()
self.simulation_sleep = simulation_sleep
self.env = None
self.env_sampler = None
self.dynamics_sample_processor = None
self.samples_data_arr = []
def construct_from_feed_dict(
self,
policy_pickle,
env_pickle,
baseline_pickle, # UNUSED
dynamics_model_pickle,
feed_dict,
):
from asynch_mb.samplers.sampler import Sampler
from asynch_mb.samplers.mb_sample_processor import ModelSampleProcessor
env = pickle.loads(env_pickle)
policy = pickle.loads(policy_pickle)
self.env = env
self.env_sampler = Sampler(env=env,
policy=policy,
**feed_dict['sampler'])
self.dynamics_sample_processor = ModelSampleProcessor(
**feed_dict['sample_processor'])
def prepare_start(self):
random_sinusoid = self.queue.get()
self.step(random_sinusoid)
self.push()
def step(self, random_sinusoid=(False, False)):
time_step = time.time()
if self.itr_counter == 1 and self.env_sampler.policy.dynamics_model.normalization is None:
if self.verbose:
logger.log('Data starts first step...')
self.env_sampler.policy.dynamics_model = pickle.loads(
self.queue.get())
if self.verbose:
logger.log('Data first step done...')
'''------------- Obtaining samples from the environment -----------'''
if self.verbose:
logger.log("Data is obtaining samples...")
env_paths = self.env_sampler.obtain_samples(
log=True,
random=random_sinusoid[0],
sinusoid=random_sinusoid[1],
log_prefix='Data-EnvSampler-',
)
'''-------------- Processing environment samples -------------------'''
if self.verbose:
logger.log("Data is processing samples...")
samples_data = self.dynamics_sample_processor.process_samples(
env_paths,
log=True,
log_prefix='Data-EnvTrajs-',
)
self.samples_data_arr.append(samples_data)
time_step = time.time() - time_step
time_sleep = max(self.simulation_sleep - time_step, 0)
time.sleep(time_sleep)
logger.logkv('Data-TimeStep', time_step)
logger.logkv('Data-TimeSleep', time_sleep)
def _synch(self, dynamics_model_state_pickle):
time_synch = time.time()
dynamics_model_state = pickle.loads(dynamics_model_state_pickle)
assert isinstance(dynamics_model_state, dict)
self.env_sampler.policy.dynamics_model.set_shared_params(
dynamics_model_state)
time_synch = time.time() - time_synch
logger.logkv('Data-TimeSynch', time_synch)
def push(self):
time_push = time.time()
self.queue_next.put(pickle.dumps(self.samples_data_arr))
self.samples_data_arr = []
time_push = time.time() - time_push
logger.logkv('Data-TimePush', time_push)
def set_stop_cond(self):
if self.itr_counter >= self.n_itr:
self.stop_cond.set()