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()
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()
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': {
'step_size': kwargs['step_size'],
}
}
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,
)
trainer.train()
