def run_garage(env, seed, log_dir):
"""
Create garage model and training.
Replace the ppo with the algorithm you want to run.
:param env: Environment of the task.
:param seed: Random seed for the trial.
:param log_dir: Log dir path.
:return:
"""
deterministic.set_seed(seed)
env.reset()
with LocalRunner() as runner:
env = TfEnv(env)
action_noise = OUStrategy(env.spec, sigma=params['sigma'])
policy = ContinuousMLPPolicy(
env_spec=env.spec,
hidden_sizes=params['policy_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh,
input_include_goal=True,
)
qf = ContinuousMLPQFunction(
env_spec=env.spec,
hidden_sizes=params['qf_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
input_include_goal=True,
)
replay_buffer = HerReplayBuffer(
env_spec=env.spec,
size_in_transitions=params['replay_buffer_size'],
time_horizon=params['n_rollout_steps'],
replay_k=0.4,
reward_fun=env.compute_reward,
)
algo = DDPG(
env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
policy_lr=params['policy_lr'],
qf_lr=params['qf_lr'],
plot=False,
target_update_tau=params['tau'],
n_epochs=params['n_epochs'],
n_epoch_cycles=params['n_epoch_cycles'],
n_train_steps=params['n_train_steps'],
discount=params['discount'],
exploration_strategy=action_noise,
policy_optimizer=tf.train.AdamOptimizer,
qf_optimizer=tf.train.AdamOptimizer,
buffer_batch_size=256,
input_include_goal=True,
)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
logger.add_output(StdOutput())
logger.add_output(CsvOutput(tabular_log_file))
logger.add_output(TensorBoardOutput(log_dir))
runner.setup(algo, env)
runner.train(
n_epochs=params['n_epochs'],
n_epoch_cycles=params['n_epoch_cycles'],
batch_size=params['n_rollout_steps'])
logger.remove_all()
return tabular_log_file
def pearl_half_cheetah_vel(ctxt=None,
seed=1,
num_epochs=500,
num_train_tasks=100,
num_test_tasks=30,
latent_size=5,
encoder_hidden_size=200,
net_size=300,
meta_batch_size=16,
num_steps_per_epoch=2000,
num_initial_steps=2000,
num_tasks_sample=5,
num_steps_prior=400,
num_extra_rl_steps_posterior=600,
batch_size=256,
embedding_batch_size=100,
embedding_mini_batch_size=100,
max_path_length=200,
reward_scale=5.,
use_gpu=False):
"""Train PEARL with HalfCheetahVel environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
num_epochs (int): Number of training epochs.
num_train_tasks (int): Number of tasks for training.
num_test_tasks (int): Number of tasks for testing.
latent_size (int): Size of latent context vector.
encoder_hidden_size (int): Output dimension of dense layer of the
context encoder.
net_size (int): Output dimension of a dense layer of Q-function and
value function.
meta_batch_size (int): Meta batch size.
num_steps_per_epoch (int): Number of iterations per epoch.
num_initial_steps (int): Number of transitions obtained per task before
training.
num_tasks_sample (int): Number of random tasks to obtain data for each
iteration.
num_steps_prior (int): Number of transitions to obtain per task with
z ~ prior.
num_extra_rl_steps_posterior (int): Number of additional transitions
to obtain per task with z ~ posterior that are only used to train
the policy and NOT the encoder.
batch_size (int): Number of transitions in RL batch.
embedding_batch_size (int): Number of transitions in context batch.
embedding_mini_batch_size (int): Number of transitions in mini context
batch; should be same as embedding_batch_size for non-recurrent
encoder.
max_path_length (int): Maximum path length.
reward_scale (int): Reward scale.
use_gpu (bool): Whether or not to use GPU for training.
"""
set_seed(seed)
encoder_hidden_sizes = (encoder_hidden_size, encoder_hidden_size,
encoder_hidden_size)
# create multi-task environment and sample tasks
env_sampler = SetTaskSampler(
lambda: GarageEnv(normalize(HalfCheetahVelEnv())))
env = env_sampler.sample(num_train_tasks)
test_env_sampler = SetTaskSampler(
lambda: GarageEnv(normalize(HalfCheetahVelEnv())))
runner = LocalRunner(ctxt)
# instantiate networks
augmented_env = PEARL.augment_env_spec(env[0](), latent_size)
qf = ContinuousMLPQFunction(env_spec=augmented_env,
hidden_sizes=[net_size, net_size, net_size])
vf_env = PEARL.get_env_spec(env[0](), latent_size, 'vf')
vf = ContinuousMLPQFunction(env_spec=vf_env,
hidden_sizes=[net_size, net_size, net_size])
inner_policy = TanhGaussianMLPPolicy(
env_spec=augmented_env, hidden_sizes=[net_size, net_size, net_size])
pearl = PEARL(
env=env,
policy_class=ContextConditionedPolicy,
encoder_class=MLPEncoder,
inner_policy=inner_policy,
qf=qf,
vf=vf,
num_train_tasks=num_train_tasks,
num_test_tasks=num_test_tasks,
latent_dim=latent_size,
encoder_hidden_sizes=encoder_hidden_sizes,
test_env_sampler=test_env_sampler,
meta_batch_size=meta_batch_size,
num_steps_per_epoch=num_steps_per_epoch,
num_initial_steps=num_initial_steps,
num_tasks_sample=num_tasks_sample,
num_steps_prior=num_steps_prior,
num_extra_rl_steps_posterior=num_extra_rl_steps_posterior,
batch_size=batch_size,
embedding_batch_size=embedding_batch_size,
embedding_mini_batch_size=embedding_mini_batch_size,
max_path_length=max_path_length,
reward_scale=reward_scale,
)
tu.set_gpu_mode(use_gpu, gpu_id=0)
if use_gpu:
pearl.to()
runner.setup(algo=pearl,
env=env[0](),
sampler_cls=LocalSampler,
sampler_args=dict(max_path_length=max_path_length),
n_workers=1,
worker_class=PEARLWorker)
runner.train(n_epochs=num_epochs, batch_size=batch_size)
def test_singleton_pool(self):
max_cpus = 8
with LocalRunner(max_cpus=max_cpus):
assert max_cpus == singleton_pool.n_parallel, (
'LocalRunner(max_cpu) should set up singleton_pool.')
def test_external_sess(self):
with tf.Session() as sess:
with LocalRunner(sess=sess):
pass
# sess should still be the default session here.
tf.no_op().run()
def run_task(snapshot_config, *_):
"""Set up environment and algorithm and run the task.
Args:
snapshot_config (garage.experiment.SnapshotConfig): The snapshot
configuration used by LocalRunner to create the snapshotter.
If None, it will create one with default settings.
_ : Unused parameters
"""
#count = 1
th = 1.8
g_max = 0.05
star_version = args.IS_MBPG_star
if args.env == 'CartPole':
#CartPole
env = TfEnv(normalize(CartPoleEnv()))
runner = LocalRunner(snapshot_config)
batch_size = 5000
max_length = 100
n_timestep = 5e5
n_counts = 5
name = 'CartPole'
#grad_factor = 5
grad_factor = 100
th = 1.2
# # batchsize:1
# lr = 0.1
# w = 1.5
# c = 15
#batchsize:50
lr = 0.75
c = 1
w = 1
# for MBPG+:
# lr = 1.2
#g_max = 0.03
discount = 0.995
path = './init/CartPole_policy.pth'
if args.env == 'Walker':
#Walker_2d
env = TfEnv(normalize(Walker2dEnv()))
runner = LocalRunner(snapshot_config)
batch_size = 50000
max_length = 500
th = 1.2
n_timestep = 1e7
n_counts = 5
lr = 0.75
w = 2
c = 5
grad_factor = 10
# for MBPG+:
#lr = 0.9
discount = 0.999
name = 'Walk'
path = './init/Walk_policy.pth'
if args.env == 'Hopper':
#Hopper
env = TfEnv(normalize(HopperEnv()))
runner = LocalRunner(snapshot_config)
batch_size = 50000
max_length = 1000
th = 1.5
n_timestep = 1e7
n_counts = 5
lr = 0.75
w = 1
c = 3
grad_factor = 10
g_max = 0.15
discount = 0.999
name = 'Hopper'
path = './init/Hopper_policy.pth'
if args.env == 'HalfCheetah':
env = TfEnv(normalize(HalfCheetahEnv()))
runner = LocalRunner(snapshot_config)
batch_size = 10000
#batch_size = 50000
max_length = 500
n_timestep = 1e7
n_counts = 5
lr = 0.6
w = 3
c = 7
grad_factor = 10
th = 1.2
g_max = 0.06
discount = 0.999
name = 'HalfCheetah'
path = './init/HalfCheetah_policy.pth'
for i in range(n_counts):
print(env.spec)
if args.env == 'CartPole':
policy = CategoricalMLPPolicy(env.spec,
hidden_sizes=[8, 8],
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
else:
policy = GaussianMLPPolicy(env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
policy.load_state_dict(torch.load(path))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = MBPG_IM(
env_spec=env.spec,
env=env,
env_name=name,
policy=policy,
baseline=baseline,
max_path_length=max_length,
discount=discount,
grad_factor=grad_factor,
policy_lr=lr,
c=c,
w=w,
n_timestep=n_timestep,
#count=count,
th=th,
batch_size=batch_size,
center_adv=True,
g_max=g_max,
#decay_learning_rate=d_lr,
star_version=star_version)
runner.setup(algo, env)
runner.train(n_epochs=100, batch_size=batch_size)
def mtsac_metaworld_mt50(ctxt=None, seed=1, use_gpu=False, _gpu=0):
"""Train MTSAC with MT50 environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
use_gpu (bool): Used to enable ussage of GPU in training.
_gpu (int): The ID of the gpu (used on multi-gpu machines).
"""
deterministic.set_seed(seed)
runner = LocalRunner(ctxt)
task_names = mwb.MT50.get_train_tasks().all_task_names
train_envs = []
test_envs = []
for task_name in task_names:
train_env = normalize(GarageEnv(mwb.MT50.from_task(task_name)),
normalize_reward=True)
test_env = normalize(GarageEnv(mwb.MT50.from_task(task_name)))
train_envs.append(train_env)
test_envs.append(test_env)
mt50_train_envs = MultiEnvWrapper(train_envs,
sample_strategy=round_robin_strategy,
mode='vanilla')
mt50_test_envs = MultiEnvWrapper(test_envs,
sample_strategy=round_robin_strategy,
mode='vanilla')
policy = TanhGaussianMLPPolicy(
env_spec=mt50_train_envs.spec,
hidden_sizes=[400, 400, 400],
hidden_nonlinearity=nn.ReLU,
output_nonlinearity=None,
min_std=np.exp(-20.),
max_std=np.exp(2.),
)
qf1 = ContinuousMLPQFunction(env_spec=mt50_train_envs.spec,
hidden_sizes=[400, 400, 400],
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPQFunction(env_spec=mt50_train_envs.spec,
hidden_sizes=[400, 400, 400],
hidden_nonlinearity=F.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6), )
timesteps = 100000000
batch_size = int(150 * mt50_train_envs.num_tasks)
num_evaluation_points = 500
epochs = timesteps // batch_size
epoch_cycles = epochs // num_evaluation_points
epochs = epochs // epoch_cycles
mtsac = MTSAC(policy=policy,
qf1=qf1,
qf2=qf2,
gradient_steps_per_itr=150,
max_path_length=250,
eval_env=mt50_test_envs,
env_spec=mt50_train_envs.spec,
num_tasks=10,
steps_per_epoch=epoch_cycles,
replay_buffer=replay_buffer,
min_buffer_size=7500,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=6400)
set_gpu_mode(use_gpu, _gpu)
mtsac.to()
runner.setup(algo=mtsac, env=mt50_train_envs, sampler_cls=LocalSampler)
runner.train(n_epochs=epochs, batch_size=batch_size)
def diayn_point_mass_multigoal(ctxt=None, seed=1):
deterministic.set_seed(seed)
runner = LocalRunner(snapshot_config=ctxt)
env = MultiGoalEnv()
skills_num = 6
policy = TanhGaussianMLPSkillPolicy(
env_spec=env.spec,
skills_num=skills_num,
hidden_sizes=[256, 256],
hidden_nonlinearity=nn.ReLU,
output_nonlinearity=None,
min_std=np.exp(-20.),
max_std=np.exp(2.),
)
qf1 = ContinuousMLPSkillQFunction(env_spec=env.spec,
skills_num=skills_num,
hidden_sizes=[256, 256],
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPSkillQFunction(env_spec=env.spec,
skills_num=skills_num,
hidden_sizes=[256, 256],
hidden_nonlinearity=F.relu)
discriminator = MLPDiscriminator(env_spec=env.spec,
skills_num=skills_num,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6))
diayn = DIAYN(
env_spec=env.spec,
skills_num=skills_num,
discriminator=discriminator,
policy=policy,
qf1=qf1,
qf2=qf2,
gradient_steps_per_itr=1000,
max_path_length=500,
replay_buffer=replay_buffer,
min_buffer_size=1e4,
recorded=True, # enable the video recording func
is_gym_render=False,
media_save_path='diayn_2d_multigoal/',
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=256,
reward_scale=1.,
steps_per_epoch=1)
if torch.cuda.is_available():
tu.set_gpu_mode(True)
else:
tu.set_gpu_mode(False)
diayn.to()
worker_args = {"skills_num": skills_num}
runner.setup(algo=diayn,
env=env,
sampler_cls=LocalSkillSampler,
worker_class=SkillWorker,
worker_args=worker_args)
runner.train(n_epochs=1000, batch_size=1000)
def test_pearl_ml1_push(self):
"""Test PEARL with ML1 Push environment."""
params = dict(seed=1,
num_epochs=1,
num_train_tasks=5,
num_test_tasks=1,
latent_size=7,
encoder_hidden_sizes=[10, 10, 10],
net_size=30,
meta_batch_size=16,
num_steps_per_epoch=40,
num_initial_steps=40,
num_tasks_sample=15,
num_steps_prior=15,
num_extra_rl_steps_posterior=15,
batch_size=256,
embedding_batch_size=8,
embedding_mini_batch_size=8,
max_path_length=50,
reward_scale=10.,
use_information_bottleneck=True,
use_next_obs_in_context=False,
use_gpu=False)
net_size = params['net_size']
set_seed(params['seed'])
env_sampler = SetTaskSampler(
lambda: GarageEnv(normalize(ML1.get_train_tasks('push-v1'))))
env = env_sampler.sample(params['num_train_tasks'])
test_env_sampler = SetTaskSampler(
lambda: GarageEnv(normalize(ML1.get_test_tasks('push-v1'))))
augmented_env = PEARL.augment_env_spec(env[0](), params['latent_size'])
qf = ContinuousMLPQFunction(
env_spec=augmented_env,
hidden_sizes=[net_size, net_size, net_size])
vf_env = PEARL.get_env_spec(env[0](), params['latent_size'], 'vf')
vf = ContinuousMLPQFunction(
env_spec=vf_env, hidden_sizes=[net_size, net_size, net_size])
inner_policy = TanhGaussianMLPPolicy(
env_spec=augmented_env,
hidden_sizes=[net_size, net_size, net_size])
pearl = PEARL(
env=env,
policy_class=ContextConditionedPolicy,
encoder_class=MLPEncoder,
inner_policy=inner_policy,
qf=qf,
vf=vf,
num_train_tasks=params['num_train_tasks'],
num_test_tasks=params['num_test_tasks'],
latent_dim=params['latent_size'],
encoder_hidden_sizes=params['encoder_hidden_sizes'],
test_env_sampler=test_env_sampler,
meta_batch_size=params['meta_batch_size'],
num_steps_per_epoch=params['num_steps_per_epoch'],
num_initial_steps=params['num_initial_steps'],
num_tasks_sample=params['num_tasks_sample'],
num_steps_prior=params['num_steps_prior'],
num_extra_rl_steps_posterior=params[
'num_extra_rl_steps_posterior'],
batch_size=params['batch_size'],
embedding_batch_size=params['embedding_batch_size'],
embedding_mini_batch_size=params['embedding_mini_batch_size'],
max_path_length=params['max_path_length'],
reward_scale=params['reward_scale'],
)
set_gpu_mode(params['use_gpu'], gpu_id=0)
if params['use_gpu']:
pearl.to()
runner = LocalRunner(snapshot_config)
runner.setup(
algo=pearl,
env=env[0](),
sampler_cls=LocalSampler,
sampler_args=dict(max_path_length=params['max_path_length']),
n_workers=1,
worker_class=PEARLWorker)
runner.train(n_epochs=params['num_epochs'],
batch_size=params['batch_size'])
def kant_cheetah_hurdle(
ctxt=None,
seed=seed,
num_skills=skills_num,
num_epochs=param_num_epoches,
num_train_tasks=param_train_tasks_num,
num_test_tasks=param_test_tasks_num,
is_encoder_recurrent=False,
latent_size=param_latent_size,
encoder_hidden_size=param_encoder_hidden_size,
net_size=param_net_size,
meta_batch_size=param_meta_batch_size,
num_steps_per_epoch=param_num_steps_per_epoch,
num_initial_steps=param_num_initial_steps,
num_tasks_sample=param_num_tasks_sample,
num_steps_prior=param_num_steps_prior,
num_extra_rl_steps_posterior=param_num_extra_rl_steps_posterior,
num_skills_sample=param_num_skills_sample,
num_skills_reason_steps=param_num_skills_reason_steps,
batch_size=param_batch_size,
embedding_batch_size=param_embedding_batch_size,
embedding_mini_batch_size=param_embedding_mini_batch_size,
max_path_length=param_max_path_length,
skills_reason_reward_scale=param_skills_reason_reward_scale,
tasks_adapt_reward_scale=param_tasks_adapt_reward_scale,
use_gpu=param_use_gpu):
assert num_train_tasks is skills_num
set_seed(seed)
encoder_hidden_sizes = (encoder_hidden_size, encoder_hidden_size,
encoder_hidden_size)
ML_train_envs = [
DiaynEnvWrapper(task_proposer, skills_num, task_name,
normalize(HalfCheetahEnv_Hurdle()))
for task_name in range(skills_num)
]
env_sampler = EnvPoolSampler(ML_train_envs)
env = env_sampler.sample(num_train_tasks)
runner = LocalRunner(ctxt)
qf_env = Kant.get_env_spec(env[0](), latent_size, num_skills, "qf")
qf = ContinuousMLPQFunction(env_spec=qf_env,
hidden_sizes=[net_size, net_size, net_size])
vf_env = Kant.get_env_spec(env[0](), latent_size, num_skills, 'vf')
vf = ContinuousMLPQFunction(env_spec=vf_env,
hidden_sizes=[net_size, net_size, net_size])
controller_policy_env = Kant.get_env_spec(env[0](),
latent_size,
module="controller_policy",
num_skills=num_skills)
controller_policy = CategoricalMLPPolicy(
env_spec=controller_policy_env,
hidden_sizes=[net_size, net_size],
hidden_nonlinearity=functional.relu)
kant = Kant(
env=env,
skill_env=skill_env,
controller_policy=controller_policy,
skill_actor=skill_actor,
qf=qf,
vf=vf,
num_skills=num_skills,
num_train_tasks=num_train_tasks,
num_test_tasks=num_test_tasks,
is_encoder_recurrent=is_encoder_recurrent,
latent_dim=latent_size,
encoder_hidden_sizes=encoder_hidden_sizes,
meta_batch_size=meta_batch_size,
num_initial_steps=num_initial_steps,
num_tasks_sample=num_tasks_sample,
num_steps_per_epoch=num_steps_per_epoch,
num_steps_prior=num_steps_prior, # num_steps_posterior
num_extra_rl_steps_posterior=num_extra_rl_steps_posterior,
num_skills_reason_steps=num_skills_reason_steps,
num_skills_sample=num_skills_sample,
batch_size=batch_size,
embedding_batch_size=embedding_batch_size,
embedding_mini_batch_size=embedding_mini_batch_size,
max_path_length=max_path_length,
skills_reason_reward_scale=skills_reason_reward_scale,
tasks_adapt_reward_scale=tasks_adapt_reward_scale)
tu.set_gpu_mode(use_gpu, gpu_id=0)
if use_gpu:
kant.to()
worker_args = dict(num_skills=num_skills,
skill_actor_class=type(skill_actor),
controller_class=OpenContextConditionedControllerPolicy,
deterministic=False,
accum_context=True)
runner.setup(algo=kant,
env=env[0](),
sampler_cls=LocalSkillSampler,
sampler_args=dict(max_path_length=max_path_length),
n_workers=1,
worker_class=KantWorker,
worker_args=worker_args)
average_returns = runner.train(n_epochs=num_epochs, batch_size=batch_size)
runner.save(num_epochs - 1)
return average_returns
