def DynaMLP_get_action(self, mlp_dyna: DynamicsModel, env: Env, state,
cost_fn, num_simulated_paths, horizon):
'''
mpc.ModelBasedModelPredictiveControl.predict()
:param mlp_dyna:
:param env:
:param state:
:param cost_fn:
:param num_simulated_paths:
:param horizon:
:return:
'''
rollout = TrajectoryData(env_spec=env.env_spec)
for i in range(num_simulated_paths):
path = TransitionData(env_spec=env.env_spec)
obs = state
for j in range(horizon):
action = env.action_space.sample()
obs_ = mlp_dyna.step(action=action, state=obs)
cost = cost_fn(obs, action, obs_)
path.append(obs, action, obs_, False, -cost)
obs = obs_
rollout.append(path)
rollout.trajectories.sort(key=lambda x: x.cumulative_reward,
reverse=True)
optimial_action = rollout.trajectories[0].action_set[0]
return optimial_action
def test_trajectory_data(self):
env = make('Acrobot-v1')
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
a = TrajectoryData(env_spec)
tmp_traj = TransitionData(env_spec)
st = env.reset()
re_list = []
st_list = []
for i in range(100):
ac = env_spec.action_space.sample()
st_new, re, done, _ = env.step(action=ac)
st_list.append(st_new)
re_list.append(re)
if (i + 1) % 10 == 0:
done = True
else:
done = False
tmp_traj.append(state=st,
new_state=st_new,
action=ac,
done=done,
reward=re)
if done:
a.append(tmp_traj.get_copy())
tmp_traj.reset()
self.assertEqual(a.trajectories.__len__(), 10)
for traj in a.trajectories:
self.assertEqual(len(traj), 10)
def test_trajectory_data(self):
env = make('Acrobot-v1')
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
a = TrajectoryData(env_spec)
tmp_traj = TransitionData(env_spec)
st = env.reset()
re_list = []
st_list = []
for i in range(100):
ac = env_spec.action_space.sample()
st_new, re, done, _ = env.step(action=ac)
st_list.append(st_new)
re_list.append(re)
if (i + 1) % 10 == 0:
done = True
else:
done = False
tmp_traj.append(state=st,
new_state=st_new,
action=ac,
done=done,
reward=re)
if done is True:
a.append(tmp_traj)
tmp_traj.reset()
self.assertEqual(a.trajectories.__len__(), 10)
for traj in a.trajectories:
self.assertEqual(len(traj), 10)
data = a.return_as_transition_data()
data_gen = data.return_generator()
for d, re, st in zip(data_gen, re_list, st_list):
self.assertEqual(d[3], re)
self.assertTrue(np.equal(st, d[1]).all())
def predict(self, obs, **kwargs):
if self.is_training is True:
return self.env_spec.action_space.sample()
rollout = TrajectoryData(env_spec=self.env_spec)
state = obs
for i in range(self.parameters('SAMPLED_PATH_NUM')):
path = TransitionData(env_spec=self.env_spec)
# todo terminal_func signal problem to be consider?
for _ in range(self.parameters('SAMPLED_HORIZON')):
ac = self.policy.forward(obs=state)
new_state, re, done, _ = self.dynamics_env.step(action=ac, state=state) # step() as an Env
path.append(state=state, action=ac, new_state=new_state, reward=re, done=done)
state = new_state
rollout.append(path)
rollout.trajectories.sort(key=lambda x: x.cumulative_reward, reverse=True)
ac = rollout.trajectories[0].action_set[0]
assert self.env_spec.action_space.contains(ac)
return ac
def _sample_trajectories(self, env, agent, sample_count, init_state):
state = init_state
sample_record = TrajectoryData(self.env_spec)
done = False
for i in range(sample_count):
traj_record = TransitionData(self.env_spec)
while done is not True:
action = agent.predict(obs=state)
new_state, re, done, info = env.step(action)
if not isinstance(done, bool):
raise TypeError()
traj_record.append(state=state,
action=action,
reward=re,
new_state=new_state,
done=done)
state = new_state
state = env.reset()
sample_record.append(traj_record)
return sample_record
def predict(self, obs, is_reward_func=True):
'''
Sample SAMPLED_PATH_NUM trajectories started from 'obs'. Return the optimal action.
:param obs: Initial state.
:param reverse_sort_flag: Decide the sort direction of trajectories, set to 'True' when using reward func.
:return: Optimal action for 'obs'.
'''
rollout = TrajectoryData(env_spec=self.env_spec)
for i in range(self.parameters('SAMPLED_PATH_NUM')):
path = TransitionData(env_spec=self.env_spec)
state = obs
for j in range(self.parameters('SAMPLED_HORIZON')):
act = self.policy.forward(obs=state) # env.action_space.sample()
new_state, cost, _, _ = self.dynamics_env.step(action=act, state=state) # step() as an Env
path.append(state=state, action=act, new_state=new_state, reward=cost, done=False)
state = new_state
rollout.append(path)
rollout.trajectories.sort(key=lambda x: x.cumulative_reward, reverse=is_reward_func)
optimal_act = rollout.trajectories[0].action_set[0]
assert self.env_spec.action_space.contains(optimal_act)
return optimal_act
class PPO(ModelFreeAlgo, OnPolicyAlgo, MultiPlaceholderInput):
required_key_dict = DictConfig.load_json(
file_path=GlobalConfig().DEFAULT_PPO_REQUIRED_KEY_LIST)
@typechecked
def __init__(self,
env_spec: EnvSpec,
stochastic_policy: StochasticPolicy,
config_or_config_dict: (DictConfig, dict),
value_func: VValueFunction,
name='ppo'):
ModelFreeAlgo.__init__(self, env_spec=env_spec, name=name)
self.config = construct_dict_config(config_or_config_dict, self)
self.policy = stochastic_policy
self.value_func = value_func
to_ph_parameter_dict = dict()
self.trajectory_memory = TrajectoryData(env_spec=env_spec)
self.transition_data_for_trajectory = TransitionData(env_spec=env_spec)
self.value_func_train_data_buffer = None
# self.scaler = Scaler(obs_dim=self.env_spec.flat_obs_dim)
self.scaler = RunningStandardScaler(dims=self.env_spec.flat_obs_dim)
with tf.variable_scope(name):
self.advantages_ph = tf.placeholder(tf.float32, (None, ),
'advantages')
self.v_func_val_ph = tf.placeholder(tf.float32, (None, ),
'val_valfunc')
dist_info_list = self.policy.get_dist_info()
self.old_dist_tensor = [
(tf.placeholder(**dict(dtype=dist_info['dtype'],
shape=dist_info['shape'],
name=dist_info['name'])),
dist_info['name']) for dist_info in dist_info_list
]
self.old_policy = self.policy.make_copy(
reuse=False,
name_scope='old_{}'.format(self.policy.name),
name='old_{}'.format(self.policy.name),
distribution_tensors_tuple=tuple(self.old_dist_tensor))
to_ph_parameter_dict['beta'] = tf.placeholder(
tf.float32, (), 'beta')
to_ph_parameter_dict['eta'] = tf.placeholder(tf.float32, (), 'eta')
to_ph_parameter_dict['kl_target'] = tf.placeholder(
tf.float32, (), 'kl_target')
to_ph_parameter_dict['lr_multiplier'] = tf.placeholder(
tf.float32, (), 'lr_multiplier')
self.parameters = ParametersWithTensorflowVariable(
tf_var_list=[],
rest_parameters=dict(
advantages_ph=self.advantages_ph,
v_func_val_ph=self.v_func_val_ph,
),
to_ph_parameter_dict=to_ph_parameter_dict,
name='ppo_param',
save_rest_param_flag=False,
source_config=self.config,
require_snapshot=False)
with tf.variable_scope(name):
with tf.variable_scope('train'):
self.kl = tf.reduce_mean(self.old_policy.kl(other=self.policy))
self.policy_loss, self.policy_optimizer, self.policy_update_op = self._setup_policy_loss(
)
self.value_func_loss, self.value_func_optimizer, self.value_func_update_op = self._setup_value_func_loss(
)
var_list = get_tf_collection_var_list(
'{}/train'.format(name)) + self.policy_optimizer.variables(
) + self.value_func_optimizer.variables()
self.parameters.set_tf_var_list(
tf_var_list=sorted(list(set(var_list)), key=lambda x: x.name))
MultiPlaceholderInput.__init__(self,
sub_placeholder_input_list=[
dict(
obj=self.value_func,
attr_name='value_func',
),
dict(obj=self.policy,
attr_name='policy')
],
parameters=self.parameters)
@register_counter_info_to_status_decorator(increment=1,
info_key='init',
under_status='JUST_INITED')
def init(self, sess=None, source_obj=None):
self.policy.init()
self.value_func.init()
self.parameters.init()
if source_obj:
self.copy_from(source_obj)
super().init()
@record_return_decorator(which_recorder='self')
@register_counter_info_to_status_decorator(increment=1,
info_key='train',
under_status='TRAIN')
@typechecked
def train(self,
trajectory_data: TrajectoryData = None,
train_iter=None,
sess=None) -> dict:
super(PPO, self).train()
if trajectory_data is None:
trajectory_data = self.trajectory_memory
if len(trajectory_data) == 0:
raise MemoryBufferLessThanBatchSizeError(
'not enough trajectory data')
tf_sess = sess if sess else tf.get_default_session()
SampleProcessor.add_estimated_v_value(trajectory_data,
value_func=self.value_func)
SampleProcessor.add_discount_sum_reward(trajectory_data,
gamma=self.parameters('gamma'))
SampleProcessor.add_gae(trajectory_data,
gamma=self.parameters('gamma'),
name='advantage_set',
lam=self.parameters('lam'),
value_func=self.value_func)
train_data = trajectory_data.return_as_transition_data(
shuffle_flag=False)
SampleProcessor.normalization(train_data, key='advantage_set')
policy_res_dict = self._update_policy(
train_data=train_data,
train_iter=train_iter
if train_iter else self.parameters('policy_train_iter'),
sess=tf_sess)
value_func_res_dict = self._update_value_func(
train_data=train_data,
train_iter=train_iter
if train_iter else self.parameters('value_func_train_iter'),
sess=tf_sess)
self.trajectory_memory.reset()
return {**policy_res_dict, **value_func_res_dict}
@register_counter_info_to_status_decorator(increment=1,
info_key='test',
under_status='TEST')
def test(self, *arg, **kwargs) -> dict:
return super().test(*arg, **kwargs)
@register_counter_info_to_status_decorator(increment=1, info_key='predict')
@typechecked
def predict(self, obs: np.ndarray, sess=None, batch_flag: bool = False):
tf_sess = sess if sess else tf.get_default_session()
obs = make_batch(obs, original_shape=self.env_spec.obs_shape)
obs = self.scaler.process(data=obs)
ac = self.policy.forward(
obs=obs,
sess=tf_sess,
feed_dict=self.parameters.return_tf_parameter_feed_dict())
return ac
@typechecked
def append_to_memory(self, samples: SampleData):
# todo how to make sure the data's time sequential
iter_samples = samples.return_generator()
# scale, offset = self.scaler.get()
obs_list = []
for state, new_state, action, reward, done in iter_samples:
obs_list.append(state)
self.transition_data_for_trajectory.append(
state=self.scaler.process(state),
new_state=self.scaler.process(new_state),
action=action,
reward=reward,
done=done)
if done is True:
self.trajectory_memory.append(
self.transition_data_for_trajectory)
self.transition_data_for_trajectory.reset()
self.scaler.update_scaler(data=np.array(obs_list))
@record_return_decorator(which_recorder='self')
def save(self, global_step, save_path=None, name=None, **kwargs):
save_path = save_path if save_path else GlobalConfig(
).DEFAULT_MODEL_CHECKPOINT_PATH
name = name if name else self.name
MultiPlaceholderInput.save(self,
save_path=save_path,
global_step=global_step,
name=name,
**kwargs)
return dict(check_point_save_path=save_path,
check_point_save_global_step=global_step,
check_point_save_name=name)
@record_return_decorator(which_recorder='self')
def load(self, path_to_model, model_name, global_step=None, **kwargs):
MultiPlaceholderInput.load(self, path_to_model, model_name,
global_step, **kwargs)
return dict(check_point_load_path=path_to_model,
check_point_load_global_step=global_step,
check_point_load_name=model_name)
def _setup_policy_loss(self):
"""
Code clip from pat-cody
Three loss terms:
1) standard policy gradient
2) D_KL(pi_old || pi_new)
3) Hinge loss on [D_KL - kl_targ]^2
See: https://arxiv.org/pdf/1707.02286.pdf
"""
if self.parameters('clipping_range') is not None:
pg_ratio = tf.exp(self.policy.log_prob() -
self.old_policy.log_prob())
clipped_pg_ratio = tf.clip_by_value(
pg_ratio, 1 - self.parameters('clipping_range')[0],
1 + self.parameters('clipping_range')[1])
surrogate_loss = tf.minimum(self.advantages_ph * pg_ratio,
self.advantages_ph * clipped_pg_ratio)
loss = -tf.reduce_mean(surrogate_loss)
else:
loss1 = -tf.reduce_mean(
self.advantages_ph *
tf.exp(self.policy.log_prob() - self.old_policy.log_prob()))
loss2 = tf.reduce_mean(self.parameters('beta') * self.kl)
loss3 = self.parameters('eta') * tf.square(
tf.maximum(0.0, self.kl - 2.0 * self.parameters('kl_target')))
loss = loss1 + loss2 + loss3
self.loss1 = loss1
self.loss2 = loss2
self.loss3 = loss3
if isinstance(self.policy, PlaceholderInput):
reg_list = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES,
scope=self.policy.name_scope)
if len(reg_list) > 0:
reg_loss = tf.reduce_sum(reg_list)
loss += reg_loss
optimizer = tf.train.AdamOptimizer(
learning_rate=self.parameters('policy_lr') *
self.parameters('lr_multiplier'))
train_op = optimizer.minimize(
loss, var_list=self.policy.parameters('tf_var_list'))
return loss, optimizer, train_op
def _setup_value_func_loss(self):
# todo update the value_func design
loss = tf.reduce_mean(
tf.square(self.value_func.v_tensor - self.v_func_val_ph))
reg_loss = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES,
scope=self.value_func.name_scope)
if len(reg_loss) > 0:
loss += tf.reduce_sum(reg_loss)
optimizer = tf.train.AdamOptimizer(self.parameters('value_func_lr'))
train_op = optimizer.minimize(
loss, var_list=self.value_func.parameters('tf_var_list'))
return loss, optimizer, train_op
def _update_policy(self, train_data: TransitionData, train_iter, sess):
old_policy_feed_dict = dict()
res = sess.run(
[
getattr(self.policy, tensor[1])
for tensor in self.old_dist_tensor
],
feed_dict={
self.policy.parameters('state_input'): train_data('state_set'),
self.policy.parameters('action_input'):
train_data('action_set'),
**self.parameters.return_tf_parameter_feed_dict()
})
for tensor, val in zip(self.old_dist_tensor, res):
old_policy_feed_dict[tensor[0]] = val
feed_dict = {
self.policy.parameters('action_input'): train_data('action_set'),
self.old_policy.parameters('action_input'):
train_data('action_set'),
self.policy.parameters('state_input'): train_data('state_set'),
self.advantages_ph: train_data('advantage_set'),
**self.parameters.return_tf_parameter_feed_dict(),
**old_policy_feed_dict
}
average_loss, average_kl, average_entropy = 0.0, 0.0, 0.0
total_epoch = 0
kl = None
for i in range(train_iter):
loss, kl, entropy, _ = sess.run([
self.policy_loss, self.kl,
tf.reduce_mean(self.policy.entropy()), self.policy_update_op
],
feed_dict=feed_dict)
average_loss += loss
average_kl += kl
average_entropy += entropy
total_epoch = i + 1
if kl > self.parameters('kl_target', require_true_value=True) * 4:
# early stopping if D_KL diverges badly
break
average_loss, average_kl, average_entropy = average_loss / total_epoch, average_kl / total_epoch, average_entropy / total_epoch
if kl > self.parameters('kl_target', require_true_value=True
) * 2: # servo beta to reach D_KL target
self.parameters.set(
key='beta',
new_val=np.minimum(
35,
1.5 * self.parameters('beta', require_true_value=True)))
if self.parameters(
'beta', require_true_value=True) > 30 and self.parameters(
'lr_multiplier', require_true_value=True) > 0.1:
self.parameters.set(
key='lr_multiplier',
new_val=self.parameters('lr_multiplier',
require_true_value=True) / 1.5)
elif kl < self.parameters('kl_target', require_true_value=True) / 2:
self.parameters.set(
key='beta',
new_val=np.maximum(
1 / 35,
self.parameters('beta', require_true_value=True) / 1.5))
if self.parameters('beta', require_true_value=True) < (
1 / 30) and self.parameters('lr_multiplier',
require_true_value=True) < 10:
self.parameters.set(
key='lr_multiplier',
new_val=self.parameters('lr_multiplier',
require_true_value=True) * 1.5)
return dict(policy_average_loss=average_loss,
policy_average_kl=average_kl,
policy_average_entropy=average_entropy,
policy_total_train_epoch=total_epoch)
def _update_value_func(self, train_data: TransitionData, train_iter, sess):
if self.value_func_train_data_buffer is None:
self.value_func_train_data_buffer = train_data
else:
self.value_func_train_data_buffer.union(train_data)
y_hat = self.value_func.forward(obs=train_data('state_set'))
old_exp_var = 1 - np.var(train_data('advantage_set') - y_hat) / np.var(
train_data('advantage_set'))
for i in range(train_iter):
data_gen = self.value_func_train_data_buffer.return_generator(
batch_size=self.parameters('value_func_train_batch_size'),
infinite_run=False,
shuffle_flag=True,
assigned_keys=('state_set', 'new_state_set', 'action_set',
'reward_set', 'done_set', 'advantage_set'))
for batch in data_gen:
loss, _ = sess.run(
[self.value_func_loss, self.value_func_update_op],
feed_dict={
self.value_func.state_input: batch[0],
self.v_func_val_ph: batch[5],
**self.parameters.return_tf_parameter_feed_dict()
})
y_hat = self.value_func.forward(obs=train_data('state_set'))
loss = np.mean(np.square(y_hat - train_data('advantage_set')))
exp_var = 1 - np.var(train_data('advantage_set') - y_hat) / np.var(
train_data('advantage_set'))
self.value_func_train_data_buffer = train_data
return dict(value_func_loss=loss,
value_func_policy_exp_var=exp_var,
value_func_policy_old_exp_var=old_exp_var)
def test_init(self):
ppo, locals = self.create_ppo()
env = locals['env']
env_spec = locals['env_spec']
ppo.init()
new_ppo, _ = self.create_ppo(name='new_ppo')
new_ppo.copy_from(ppo)
self.assert_var_list_id_no_equal(
ppo.value_func.parameters('tf_var_list'),
new_ppo.value_func.parameters('tf_var_list'))
self.assert_var_list_id_no_equal(
ppo.policy.parameters('tf_var_list'),
new_ppo.policy.parameters('tf_var_list'))
self.assert_var_list_equal(
ppo.value_func.parameters('tf_var_list'),
new_ppo.value_func.parameters('tf_var_list'))
self.assert_var_list_equal(ppo.policy.parameters('tf_var_list'),
new_ppo.policy.parameters('tf_var_list'))
data = TransitionData(env_spec)
st = env.reset()
for i in range(100):
ac = ppo.predict(st)
assert ac.shape[0] == 1
self.assertTrue(env_spec.action_space.contains(ac[0]))
new_st, re, done, _ = env.step(ac)
if i % 9 == 0 and i > 0:
done = True
else:
done = False
data.append(state=st,
new_state=new_st,
action=ac,
reward=re,
done=done)
traj = TrajectoryData(env_spec=env_spec)
traj.append(data)
ppo.append_to_memory(traj)
ppo.save(save_path=GlobalConfig().DEFAULT_LOG_PATH + '/ppo_test',
global_step=0,
name=ppo.name)
for i in range(5):
ppo.append_to_memory(traj)
res = ppo.train()
print('value_func_loss {}, policy_average_loss: {}'.format(
res['value_func_loss'], res['policy_average_loss']))
traj_data = TrajectoryData(env_spec=env_spec)
traj_data.append(data)
res = ppo.train(trajectory_data=traj_data,
train_iter=5,
sess=self.sess)
print('value_func_loss {}, policy_average_loss: {}'.format(
res['value_func_loss'], res['policy_average_loss']))
self.assert_var_list_at_least_not_equal(
ppo.value_func.parameters('tf_var_list'),
new_ppo.value_func.parameters('tf_var_list'))
self.assert_var_list_at_least_not_equal(
ppo.policy.parameters('tf_var_list'),
new_ppo.policy.parameters('tf_var_list'))
ppo.load(path_to_model=GlobalConfig().DEFAULT_LOG_PATH + '/ppo_test',
model_name=ppo.name,
global_step=0)
self.assert_var_list_equal(
ppo.value_func.parameters('tf_var_list'),
new_ppo.value_func.parameters('tf_var_list'))
self.assert_var_list_equal(ppo.policy.parameters('tf_var_list'),
new_ppo.policy.parameters('tf_var_list'))