def construct_from_feed_dict(self, policy_pickle, env_pickle,
baseline_pickle, dynamics_model_pickle,
feed_dict):
from meta_mb.samplers.metrpo_samplers.metrpo_sampler import METRPOSampler
from meta_mb.samplers.bptt_samplers.bptt_sampler import BPTTSampler
from meta_mb.samplers.base import SampleProcessor
from meta_mb.algos.ppo import PPO
from meta_mb.algos.trpo import TRPO
env = pickle.loads(env_pickle)
policy = pickle.loads(policy_pickle)
baseline = pickle.loads(baseline_pickle)
dynamics_model = pickle.loads(dynamics_model_pickle)
self.policy = policy
self.baseline = baseline
if self.sampler_str == 'metrpo':
self.model_sampler = METRPOSampler(env=env,
policy=policy,
dynamics_model=dynamics_model,
**feed_dict['model_sampler'])
elif self.sampler_str == 'bptt':
self.model_sampler = BPTTSampler(env=env,
policy=policy,
dynamics_model=dynamics_model,
**feed_dict['model_sampler'])
else:
raise NotImplementedError
self.model_sample_processor = SampleProcessor(
baseline=baseline, **feed_dict['model_sample_processor'])
if self.algo == 'meppo':
self.algo = PPO(policy=policy, **feed_dict['algo'])
elif self.algo == 'metrpo':
self.algo = TRPO(policy=policy, **feed_dict['algo'])
else:
raise NotImplementedError('algo_str must be meppo or metrpo')
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()