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,
warm_up_trajectories_number=5,
use_time_index_flag=False,
name='ppo'):
ModelFreeAlgo.__init__(
self,
env_spec=env_spec,
name=name,
warm_up_trajectories_number=warm_up_trajectories_number)
self.use_time_index_flag = use_time_index_flag
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 = RunningStandardScaler(dims=self.env_spec.flat_obs_dim)
if use_time_index_flag:
scale_last_time_index_mean = self.scaler._mean
scale_last_time_index_mean[-1] = 0
scale_last_time_index_var = self.scaler._var
scale_last_time_index_var[-1] = 1000 * 1000
self.scaler.set_param(mean=scale_last_time_index_mean,
var=scale_last_time_index_var)
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_val_func')
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(self.policy))
self.average_entropy = tf.reduce_mean(self.policy.entropy())
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)
def warm_up(self, trajectory_data: TrajectoryData):
for traj in trajectory_data.trajectories:
self.scaler.update_scaler(data=traj.state_set)
if self.use_time_index_flag:
scale_last_time_index_mean = self.scaler._mean
scale_last_time_index_mean[-1] = 0
scale_last_time_index_var = self.scaler._var
scale_last_time_index_var[-1] = 1000 * 1000
self.scaler.set_param(mean=scale_last_time_index_mean,
var=scale_last_time_index_var)
@register_counter_info_to_status_decorator(increment=1,
info_key='init',
under_status='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')
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')
for i, traj in enumerate(trajectory_data.trajectories):
trajectory_data.trajectories[i].append_new_set(
name='state_set',
shape=self.env_spec.obs_shape,
data_set=np.reshape(
np.array(self.scaler.process(np.array(traj.state_set))),
[-1] + list(self.env_spec.obs_shape)))
trajectory_data.trajectories[i].append_new_set(
name='new_state_set',
shape=self.env_spec.obs_shape,
data_set=np.reshape(
np.array(self.scaler.process(np.array(
traj.new_state_set))),
[-1] + list(self.env_spec.obs_shape)))
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)
trajectory_data = SampleProcessor.normalization(trajectory_data,
key='advantage_set')
policy_res_dict = self._update_policy(
state_set=np.concatenate(
[t('state_set') for t in trajectory_data.trajectories],
axis=0),
action_set=np.concatenate(
[t('action_set') for t in trajectory_data.trajectories],
axis=0),
advantage_set=np.concatenate(
[t('advantage_set') for t in trajectory_data.trajectories],
axis=0),
train_iter=train_iter
if train_iter else self.parameters('policy_train_iter'),
sess=tf_sess)
value_func_res_dict = self._update_value_func(
state_set=np.concatenate(
[t('state_set') for t in trajectory_data.trajectories],
axis=0),
discount_set=np.concatenate(
[t('discount_set') for t in trajectory_data.trajectories],
axis=0),
train_iter=train_iter
if train_iter else self.parameters('value_func_train_iter'),
sess=tf_sess)
trajectory_data.reset()
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')
def predict(self, obs: np.ndarray, sess=None, batch_flag: bool = False):
tf_sess = sess if sess else tf.get_default_session()
ac = self.policy.forward(
obs=self.scaler.process(
data=make_batch(obs, original_shape=self.env_spec.obs_shape)),
sess=tf_sess,
feed_dict=self.parameters.return_tf_parameter_feed_dict())
return ac
def append_to_memory(self, samples: TrajectoryData):
# todo how to make sure the data's time sequential
obs_list = samples.trajectories[0].state_set
for i in range(1, len(samples.trajectories)):
obs_list = np.array(
np.concatenate([obs_list, samples.trajectories[i].state_set],
axis=0))
self.trajectory_memory.union(samples)
self.scaler.update_scaler(data=np.array(obs_list))
if self.use_time_index_flag:
scale_last_time_index_mean = self.scaler._mean
scale_last_time_index_mean[-1] = 0
scale_last_time_index_var = self.scaler._var
scale_last_time_index_var[-1] = 1000 * 1000
self.scaler.set_param(mean=scale_last_time_index_mean,
var=scale_last_time_index_var)
@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(
tf.squeeze(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, state_set, action_set, advantage_set, 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'): state_set,
self.policy.parameters('action_input'): 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'): action_set,
self.old_policy.parameters('action_input'): action_set,
self.policy.parameters('state_input'): state_set,
self.advantages_ph: 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):
_ = sess.run(self.policy_update_op, feed_dict=feed_dict)
loss, kl, entropy = sess.run(
[self.policy_loss, self.kl, self.average_entropy],
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, state_set, discount_set, train_iter, sess):
y_hat = self.value_func.forward(obs=state_set).squeeze()
old_exp_var = 1 - np.var(discount_set - y_hat) / np.var(discount_set)
if self.value_func_train_data_buffer is None:
self.value_func_train_data_buffer = (state_set, discount_set)
else:
self.value_func_train_data_buffer = (
np.concatenate(
[self.value_func_train_data_buffer[0], state_set], axis=0),
np.concatenate(
[self.value_func_train_data_buffer[1], discount_set],
axis=0))
if len(self.value_func_train_data_buffer[0]) > self.parameters(
'value_func_memory_size'):
self.value_func_train_data_buffer = tuple(
np.array(data[-self.parameters('value_func_memory_size'):])
for data in self.value_func_train_data_buffer)
state_set_all, discount_set_all = self.value_func_train_data_buffer
param_dict = self.parameters.return_tf_parameter_feed_dict()
for i in range(train_iter):
random_index = np.random.choice(np.arange(len(state_set_all)),
len(state_set_all))
state_set_all = state_set_all[random_index]
discount_set_all = discount_set_all[random_index]
for index in range(
0,
len(state_set_all) -
self.parameters('value_func_train_batch_size'),
self.parameters('value_func_train_batch_size')):
state = np.array(
state_set_all[index:index + self.
parameters('value_func_train_batch_size')])
discount = discount_set_all[
index:index +
self.parameters('value_func_train_batch_size')]
loss, _ = sess.run(
[self.value_func_loss, self.value_func_update_op],
options=tf.RunOptions(
report_tensor_allocations_upon_oom=True),
feed_dict={
self.value_func.state_input: state,
self.v_func_val_ph: discount,
**param_dict
})
y_hat = self.value_func.forward(obs=state_set).squeeze()
loss = np.mean(np.square(y_hat - discount_set))
exp_var = 1 - np.var(discount_set - y_hat) / np.var(discount_set)
return dict(value_func_loss=loss,
value_func_policy_exp_var=exp_var,
value_func_policy_old_exp_var=old_exp_var)
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')
for i, traj in enumerate(trajectory_data.trajectories):
trajectory_data.trajectories[i].append_new_set(
name='state_set',
shape=self.env_spec.obs_shape,
data_set=np.reshape(
np.array(self.scaler.process(np.array(traj.state_set))),
[-1] + list(self.env_spec.obs_shape)))
trajectory_data.trajectories[i].append_new_set(
name='new_state_set',
shape=self.env_spec.obs_shape,
data_set=np.reshape(
np.array(self.scaler.process(np.array(
traj.new_state_set))),
[-1] + list(self.env_spec.obs_shape)))
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)
trajectory_data = SampleProcessor.normalization(trajectory_data,
key='advantage_set')
policy_res_dict = self._update_policy(
state_set=np.concatenate(
[t('state_set') for t in trajectory_data.trajectories],
axis=0),
action_set=np.concatenate(
[t('action_set') for t in trajectory_data.trajectories],
axis=0),
advantage_set=np.concatenate(
[t('advantage_set') for t in trajectory_data.trajectories],
axis=0),
train_iter=train_iter
if train_iter else self.parameters('policy_train_iter'),
sess=tf_sess)
value_func_res_dict = self._update_value_func(
state_set=np.concatenate(
[t('state_set') for t in trajectory_data.trajectories],
axis=0),
discount_set=np.concatenate(
[t('discount_set') for t in trajectory_data.trajectories],
axis=0),
train_iter=train_iter
if train_iter else self.parameters('value_func_train_iter'),
sess=tf_sess)
trajectory_data.reset()
self.trajectory_memory.reset()
return {**policy_res_dict, **value_func_res_dict}