def learn(self, total_timesteps, callback=None, log_interval=100, tb_log_name="A2C",
reset_num_timesteps=True):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
callback = self._init_callback(callback)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn()
self.learning_rate_schedule = Scheduler(initial_value=self.learning_rate, n_values=total_timesteps,
schedule=self.lr_schedule)
t_start = time.time()
callback.on_training_start(locals(), globals())
for update in range(1, total_timesteps // self.n_batch + 1):
callback.on_rollout_start()
# true_reward is the reward without discount
rollout = self.runner.run(callback)
# unpack
obs, states, rewards, masks, actions, values, ep_infos, true_reward = rollout
callback.update_locals(locals())
callback.on_rollout_end()
# Early stopping due to the callback
if not self.runner.continue_training:
break
self.ep_info_buf.extend(ep_infos)
_, value_loss, policy_entropy = self._train_step(obs, states, rewards, masks, actions, values,
self.num_timesteps // self.n_batch, writer)
n_seconds = time.time() - t_start
fps = int((update * self.n_batch) / n_seconds)
if writer is not None:
total_episode_reward_logger(self.episode_reward,
true_reward.reshape((self.n_envs, self.n_steps)),
masks.reshape((self.n_envs, self.n_steps)),
writer, self.num_timesteps)
if self.verbose >= 1 and (update % log_interval == 0 or update == 1):
explained_var = explained_variance(values, rewards)
logger.record_tabular("nupdates", update)
logger.record_tabular("total_timesteps", self.num_timesteps)
logger.record_tabular("fps", fps)
logger.record_tabular("policy_entropy", float(policy_entropy))
logger.record_tabular("value_loss", float(value_loss))
logger.record_tabular("explained_variance", float(explained_var))
if len(self.ep_info_buf) > 0 and len(self.ep_info_buf[0]) > 0:
logger.logkv('ep_reward_mean', safe_mean([ep_info['r'] for ep_info in self.ep_info_buf]))
logger.logkv('ep_len_mean', safe_mean([ep_info['l'] for ep_info in self.ep_info_buf]))
logger.dump_tabular()
callback.on_training_end()
return self
def setup_model(self):
with SetVerbosity(self.verbose):
assert not isinstance(self.action_space, gym.spaces.Box), \
"Error: DQN cannot output a gym.spaces.Box action space."
# If the policy is wrap in functool.partial (e.g. to disable dueling)
# unwrap it to check the class type
if isinstance(self.policy, partial):
test_policy = self.policy.func
else:
test_policy = self.policy
assert issubclass(test_policy, DQNPolicy), "Error: the input policy for the DQN model must be " \
"an instance of DQNPolicy."
self.graph = tf.Graph()
with self.graph.as_default():
self.set_random_seed(self.seed)
self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess,
graph=self.graph)
optimizer = tf.train.AdamOptimizer(
learning_rate=self.learning_rate)
self.act, self._train_step, self.update_target, self.step_model = build_train(
q_func=partial(self.policy, **self.policy_kwargs),
ob_space=self.observation_space,
ac_space=self.action_space,
optimizer=optimizer,
gamma=self.gamma,
grad_norm_clipping=10,
param_noise=self.param_noise,
sess=self.sess,
full_tensorboard_log=self.full_tensorboard_log,
double_q=self.double_q)
self.proba_step = self.step_model.proba_step
self.params = tf_util.get_trainable_vars("deepq")
# Initialize the parameters and copy them to the target network.
tf_util.initialize(self.sess)
self.update_target(sess=self.sess)
self.summary = tf.summary.merge_all()
def setup_model(self):
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the A2C model must be an " \
"instance of common.policies.ActorCriticPolicy."
self.graph = tf.Graph()
with self.graph.as_default():
self.set_random_seed(self.seed)
self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess,
graph=self.graph)
self.n_batch = self.n_envs * self.n_steps
n_batch_step = None
n_batch_train = None
if issubclass(self.policy, RecurrentActorCriticPolicy):
n_batch_step = self.n_envs
n_batch_train = self.n_envs * self.n_steps
step_model = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
n_batch_step,
reuse=False,
**self.policy_kwargs)
with tf.variable_scope(
"train_model",
reuse=True,
custom_getter=tf_util.outer_scope_getter(
"train_model")):
train_model = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
self.n_steps,
n_batch_train,
reuse=True,
**self.policy_kwargs)
with tf.variable_scope("loss", reuse=False):
self.actions_ph = train_model.pdtype.sample_placeholder(
[None], name="action_ph")
self.advs_ph = tf.placeholder(tf.float32, [None],
name="advs_ph")
self.rewards_ph = tf.placeholder(tf.float32, [None],
name="rewards_ph")
self.learning_rate_ph = tf.placeholder(
tf.float32, [], name="learning_rate_ph")
neglogpac = train_model.proba_distribution.neglogp(
self.actions_ph)
self.entropy = tf.reduce_mean(
train_model.proba_distribution.entropy())
self.pg_loss = tf.reduce_mean(self.advs_ph * neglogpac)
self.vf_loss = mse(tf.squeeze(train_model.value_flat),
self.rewards_ph)
# https://arxiv.org/pdf/1708.04782.pdf#page=9, https://arxiv.org/pdf/1602.01783.pdf#page=4
# and https://github.com/dennybritz/reinforcement-learning/issues/34
# suggest to add an entropy component in order to improve exploration.
loss = self.pg_loss - self.entropy * self.ent_coef + self.vf_loss * self.vf_coef
tf.summary.scalar('entropy_loss', self.entropy)
tf.summary.scalar('policy_gradient_loss', self.pg_loss)
tf.summary.scalar('value_function_loss', self.vf_loss)
tf.summary.scalar('loss', loss)
self.params = tf_util.get_trainable_vars("model")
grads = tf.gradients(loss, self.params)
if self.max_grad_norm is not None:
grads, _ = tf.clip_by_global_norm(
grads, self.max_grad_norm)
grads = list(zip(grads, self.params))
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards',
tf.reduce_mean(self.rewards_ph))
tf.summary.scalar('learning_rate',
tf.reduce_mean(self.learning_rate_ph))
tf.summary.scalar('advantage',
tf.reduce_mean(self.advs_ph))
if self.full_tensorboard_log:
tf.summary.histogram('discounted_rewards',
self.rewards_ph)
tf.summary.histogram('learning_rate',
self.learning_rate_ph)
tf.summary.histogram('advantage', self.advs_ph)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', train_model.obs_ph)
else:
tf.summary.histogram('observation',
train_model.obs_ph)
trainer = tf.train.RMSPropOptimizer(
learning_rate=self.learning_rate_ph,
decay=self.alpha,
epsilon=self.epsilon,
momentum=self.momentum)
self.apply_backprop = trainer.apply_gradients(grads)
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.proba_step = step_model.proba_step
self.value = step_model.value
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=self.sess)
self.summary = tf.summary.merge_all()
def learn(self,
total_timesteps,
callback=None,
log_interval=100,
tb_log_name="ACKTR",
reset_num_timesteps=True):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
callback = self._init_callback(callback)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn()
self.n_batch = self.n_envs * self.n_steps
self.learning_rate_schedule = Scheduler(
initial_value=self.learning_rate,
n_values=total_timesteps,
schedule=self.lr_schedule)
# FIFO queue of the q_runner thread is closed at the end of the learn function.
# As a result, it needs to be redefinied at every call
with self.graph.as_default():
with tf.variable_scope(
"kfac_apply",
reuse=self.trained,
custom_getter=tf_util.outer_scope_getter(
"kfac_apply")):
# Some of the variables are not in a scope when they are create
# so we make a note of any previously uninitialized variables
tf_vars = tf.global_variables()
is_uninitialized = self.sess.run(
[tf.is_variable_initialized(var) for var in tf_vars])
old_uninitialized_vars = [
v for (v, f) in zip(tf_vars, is_uninitialized) if not f
]
self.train_op, self.q_runner = self.optim.apply_gradients(
list(zip(self.grads_check, self.params)))
# then we check for new uninitialized variables and initialize them
tf_vars = tf.global_variables()
is_uninitialized = self.sess.run(
[tf.is_variable_initialized(var) for var in tf_vars])
new_uninitialized_vars = [
v for (v, f) in zip(tf_vars, is_uninitialized)
if not f and v not in old_uninitialized_vars
]
if len(new_uninitialized_vars) != 0:
self.sess.run(
tf.variables_initializer(new_uninitialized_vars))
self.trained = True
t_start = time.time()
coord = tf.train.Coordinator()
if self.q_runner is not None:
enqueue_threads = self.q_runner.create_threads(self.sess,
coord=coord,
start=True)
else:
enqueue_threads = []
callback.on_training_start(locals(), globals())
for update in range(1, total_timesteps // self.n_batch + 1):
callback.on_rollout_start()
# pytype:disable=bad-unpacking
# true_reward is the reward without discount
if isinstance(self.runner, PPO2Runner):
# We are using GAE
rollout = self.runner.run(callback)
obs, returns, masks, actions, values, _, states, ep_infos, true_reward = rollout
else:
rollout = self.runner.run(callback)
obs, states, returns, masks, actions, values, ep_infos, true_reward = rollout
# pytype:enable=bad-unpacking
callback.update_locals(locals())
callback.on_rollout_end()
# Early stopping due to the callback
if not self.runner.continue_training:
break
self.ep_info_buf.extend(ep_infos)
policy_loss, value_loss, policy_entropy = self._train_step(
obs, states, returns, masks, actions, values,
self.num_timesteps // (self.n_batch + 1), writer)
n_seconds = time.time() - t_start
fps = int((update * self.n_batch) / n_seconds)
if writer is not None:
total_episode_reward_logger(
self.episode_reward,
true_reward.reshape((self.n_envs, self.n_steps)),
masks.reshape((self.n_envs, self.n_steps)), writer,
self.num_timesteps)
if self.verbose >= 1 and (update % log_interval == 0
or update == 1):
explained_var = explained_variance(values, returns)
logger.record_tabular("nupdates", update)
logger.record_tabular("total_timesteps",
self.num_timesteps)
logger.record_tabular("fps", fps)
logger.record_tabular("policy_entropy",
float(policy_entropy))
logger.record_tabular("policy_loss", float(policy_loss))
logger.record_tabular("value_loss", float(value_loss))
logger.record_tabular("explained_variance",
float(explained_var))
if len(self.ep_info_buf) > 0 and len(
self.ep_info_buf[0]) > 0:
logger.logkv(
'ep_reward_mean',
safe_mean([
ep_info['r'] for ep_info in self.ep_info_buf
]))
logger.logkv(
'ep_len_mean',
safe_mean([
ep_info['l'] for ep_info in self.ep_info_buf
]))
logger.dump_tabular()
coord.request_stop()
coord.join(enqueue_threads)
callback.on_training_end()
return self
def setup_model(self):
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the ACKTR model must be " \
"an instance of common.policies.ActorCriticPolicy."
# Enable continuous actions tricks (normalized advantage)
self.continuous_actions = isinstance(self.action_space, Box)
self.graph = tf.Graph()
with self.graph.as_default():
self.set_random_seed(self.seed)
self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess,
graph=self.graph)
n_batch_step = None
n_batch_train = None
if issubclass(self.policy, RecurrentActorCriticPolicy):
n_batch_step = self.n_envs
n_batch_train = self.n_envs * self.n_steps
step_model = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
n_batch_step,
reuse=False,
**self.policy_kwargs)
self.params = params = tf_util.get_trainable_vars("model")
with tf.variable_scope(
"train_model",
reuse=True,
custom_getter=tf_util.outer_scope_getter(
"train_model")):
train_model = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
self.n_steps,
n_batch_train,
reuse=True,
**self.policy_kwargs)
with tf.variable_scope(
"loss",
reuse=False,
custom_getter=tf_util.outer_scope_getter("loss")):
self.advs_ph = advs_ph = tf.placeholder(tf.float32, [None])
self.rewards_ph = rewards_ph = tf.placeholder(
tf.float32, [None])
self.learning_rate_ph = learning_rate_ph = tf.placeholder(
tf.float32, [])
self.actions_ph = train_model.pdtype.sample_placeholder(
[None])
neg_log_prob = train_model.proba_distribution.neglogp(
self.actions_ph)
# training loss
pg_loss = tf.reduce_mean(advs_ph * neg_log_prob)
self.entropy = entropy = tf.reduce_mean(
train_model.proba_distribution.entropy())
self.pg_loss = pg_loss = pg_loss - self.ent_coef * entropy
self.vf_loss = vf_loss = mse(
tf.squeeze(train_model.value_fn), rewards_ph)
train_loss = pg_loss + self.vf_coef * vf_loss
# Fisher loss construction
self.pg_fisher = pg_fisher_loss = -tf.reduce_mean(
neg_log_prob)
sample_net = train_model.value_fn + tf.random_normal(
tf.shape(train_model.value_fn))
self.vf_fisher = vf_fisher_loss = -self.vf_fisher_coef * tf.reduce_mean(
tf.pow(
train_model.value_fn -
tf.stop_gradient(sample_net), 2))
self.joint_fisher = pg_fisher_loss + vf_fisher_loss
tf.summary.scalar('entropy_loss', self.entropy)
tf.summary.scalar('policy_gradient_loss', pg_loss)
tf.summary.scalar('policy_gradient_fisher_loss',
pg_fisher_loss)
tf.summary.scalar('value_function_loss', self.vf_loss)
tf.summary.scalar('value_function_fisher_loss',
vf_fisher_loss)
tf.summary.scalar('loss', train_loss)
self.grads_check = tf.gradients(train_loss, params)
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards',
tf.reduce_mean(self.rewards_ph))
tf.summary.scalar('learning_rate',
tf.reduce_mean(self.learning_rate_ph))
tf.summary.scalar('advantage',
tf.reduce_mean(self.advs_ph))
if self.full_tensorboard_log:
tf.summary.histogram('discounted_rewards',
self.rewards_ph)
tf.summary.histogram('learning_rate',
self.learning_rate_ph)
tf.summary.histogram('advantage', self.advs_ph)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', train_model.obs_ph)
else:
tf.summary.histogram('observation',
train_model.obs_ph)
with tf.variable_scope(
"kfac",
reuse=False,
custom_getter=tf_util.outer_scope_getter("kfac")):
with tf.device('/gpu:0'):
self.optim = optim = kfac.KfacOptimizer(
learning_rate=learning_rate_ph,
clip_kl=self.kfac_clip,
momentum=0.9,
kfac_update=self.kfac_update,
epsilon=0.01,
stats_decay=0.99,
async_eigen_decomp=self.async_eigen_decomp,
cold_iter=10,
max_grad_norm=self.max_grad_norm,
verbose=self.verbose)
optim.compute_and_apply_stats(self.joint_fisher,
var_list=params)
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.proba_step = step_model.proba_step
self.value = step_model.value
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=self.sess)
self.summary = tf.summary.merge_all()
def learn(self,
total_timesteps,
callback=None,
log_interval=100,
tb_log_name="DQN",
reset_num_timesteps=True,
replay_wrapper=None):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
callback = self._init_callback(callback)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn()
# Create the replay buffer
if self.prioritized_replay:
self.replay_buffer = PrioritizedReplayBuffer(
self.buffer_size, alpha=self.prioritized_replay_alpha)
if self.prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
else:
prioritized_replay_beta_iters = self.prioritized_replay_beta_iters
self.beta_schedule = LinearSchedule(
prioritized_replay_beta_iters,
initial_p=self.prioritized_replay_beta0,
final_p=1.0)
else:
self.replay_buffer = ReplayBuffer(self.buffer_size)
self.beta_schedule = None
if replay_wrapper is not None:
assert not self.prioritized_replay, "Prioritized replay buffer is not supported by HER"
self.replay_buffer = replay_wrapper(self.replay_buffer)
# Create the schedule for exploration starting from 1.
self.exploration = LinearSchedule(
schedule_timesteps=int(self.exploration_fraction *
total_timesteps),
initial_p=self.exploration_initial_eps,
final_p=self.exploration_final_eps)
episode_rewards = [0.0]
episode_successes = []
callback.on_training_start(locals(), globals())
callback.on_rollout_start()
reset = True
obs = self.env.reset()
# Retrieve unnormalized observation for saving into the buffer
if self._vec_normalize_env is not None:
obs_ = self._vec_normalize_env.get_original_obs().squeeze()
for _ in range(total_timesteps):
# Take action and update exploration to the newest value
kwargs = {}
if not self.param_noise:
update_eps = self.exploration.value(self.num_timesteps)
update_param_noise_threshold = 0.
else:
update_eps = 0.
# Compute the threshold such that the KL divergence between perturbed and non-perturbed
# policy is comparable to eps-greedy exploration with eps = exploration.value(t).
# See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017
# for detailed explanation.
update_param_noise_threshold = \
-np.log(1. - self.exploration.value(self.num_timesteps) +
self.exploration.value(self.num_timesteps) / float(self.env.action_space.n))
kwargs['reset'] = reset
kwargs[
'update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
with self.sess.as_default():
action = self.act(np.array(obs)[None],
update_eps=update_eps,
**kwargs)[0]
env_action = action
reset = False
new_obs, rew, done, info = self.env.step(env_action)
self.num_timesteps += 1
# Stop training if return value is False
callback.update_locals(locals())
if callback.on_step() is False:
break
# Store only the unnormalized version
if self._vec_normalize_env is not None:
new_obs_ = self._vec_normalize_env.get_original_obs(
).squeeze()
reward_ = self._vec_normalize_env.get_original_reward(
).squeeze()
else:
# Avoid changing the original ones
obs_, new_obs_, reward_ = obs, new_obs, rew
# Store transition in the replay buffer.
self.replay_buffer_add(obs_, action, reward_, new_obs_, done,
info)
obs = new_obs
# Save the unnormalized observation
if self._vec_normalize_env is not None:
obs_ = new_obs_
if writer is not None:
ep_rew = np.array([reward_]).reshape((1, -1))
ep_done = np.array([done]).reshape((1, -1))
tf_util.total_episode_reward_logger(
self.episode_reward, ep_rew, ep_done, writer,
self.num_timesteps)
episode_rewards[-1] += reward_
if done:
maybe_is_success = info.get('is_success')
if maybe_is_success is not None:
episode_successes.append(float(maybe_is_success))
if not isinstance(self.env, VecEnv):
obs = self.env.reset()
episode_rewards.append(0.0)
reset = True
# Do not train if the warmup phase is not over
# or if there are not enough samples in the replay buffer
can_sample = self.replay_buffer.can_sample(self.batch_size)
if can_sample and self.num_timesteps > self.learning_starts \
and self.num_timesteps % self.train_freq == 0:
callback.on_rollout_end()
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
# pytype:disable=bad-unpacking
if self.prioritized_replay:
assert self.beta_schedule is not None, \
"BUG: should be LinearSchedule when self.prioritized_replay True"
experience = self.replay_buffer.sample(
self.batch_size,
beta=self.beta_schedule.value(self.num_timesteps),
env=self._vec_normalize_env)
(obses_t, actions, rewards, obses_tp1, dones, weights,
batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample(
self.batch_size, env=self._vec_normalize_env)
weights, batch_idxes = np.ones_like(rewards), None
# pytype:enable=bad-unpacking
if writer is not None:
# run loss backprop with summary, but once every 100 steps save the metadata
# (memory, compute time, ...)
if (1 + self.num_timesteps) % 100 == 0:
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, td_errors = self._train_step(
obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
writer.add_run_metadata(
run_metadata, 'step%d' % self.num_timesteps)
else:
summary, td_errors = self._train_step(
obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess)
writer.add_summary(summary, self.num_timesteps)
else:
_, td_errors = self._train_step(obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess)
if self.prioritized_replay:
new_priorities = np.abs(
td_errors) + self.prioritized_replay_eps
assert isinstance(self.replay_buffer,
PrioritizedReplayBuffer)
self.replay_buffer.update_priorities(
batch_idxes, new_priorities)
callback.on_rollout_start()
if can_sample and self.num_timesteps > self.learning_starts and \
self.num_timesteps % self.target_network_update_freq == 0:
# Update target network periodically.
self.update_target(sess=self.sess)
if len(episode_rewards[-101:-1]) == 0:
mean_100ep_reward = -np.inf
else:
mean_100ep_reward = round(
float(np.mean(episode_rewards[-101:-1])), 1)
num_episodes = len(episode_rewards)
if self.verbose >= 1 and done and log_interval is not None and len(
episode_rewards) % log_interval == 0:
logger.record_tabular("steps", self.num_timesteps)
logger.record_tabular("episodes", num_episodes)
if len(episode_successes) > 0:
logger.logkv("success rate",
np.mean(episode_successes[-100:]))
logger.record_tabular("mean 100 episode reward",
mean_100ep_reward)
logger.record_tabular(
"% time spent exploring",
int(100 * self.exploration.value(self.num_timesteps)))
logger.dump_tabular()
callback.on_training_end()
return self
def learn(self,
total_timesteps,
callback=None,
log_interval=1,
tb_log_name="PPO2",
reset_num_timesteps=True):
# Transform to callable if needed
self.learning_rate = get_schedule_fn(self.learning_rate)
self.cliprange = get_schedule_fn(self.cliprange)
cliprange_vf = get_schedule_fn(self.cliprange_vf)
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
callback = self._init_callback(callback)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn()
t_first_start = time.time()
n_updates = total_timesteps // self.n_batch
callback.on_training_start(locals(), globals())
for update in range(1, n_updates + 1):
assert self.n_batch % self.nminibatches == 0, (
"The number of minibatches (`nminibatches`) "
"is not a factor of the total number of samples "
"collected per rollout (`n_batch`), "
"some samples won't be used.")
batch_size = self.n_batch // self.nminibatches
t_start = time.time()
frac = 1.0 - (update - 1.0) / n_updates
lr_now = self.learning_rate(frac)
cliprange_now = self.cliprange(frac)
cliprange_vf_now = cliprange_vf(frac)
callback.on_rollout_start()
# true_reward is the reward without discount
rollout = self.runner.run(callback)
# Unpack
obs, returns, masks, actions, values, neglogpacs, states, ep_infos, true_reward = rollout
callback.on_rollout_end()
# Early stopping due to the callback
if not self.runner.continue_training:
break
self.ep_info_buf.extend(ep_infos)
mb_loss_vals = []
if states is None: # nonrecurrent version
update_fac = max(
self.n_batch // self.nminibatches // self.noptepochs,
1)
inds = np.arange(self.n_batch)
for epoch_num in range(self.noptepochs):
np.random.shuffle(inds)
for start in range(0, self.n_batch, batch_size):
timestep = self.num_timesteps // update_fac + (
(epoch_num * self.n_batch + start) //
batch_size)
end = start + batch_size
mbinds = inds[start:end]
slices = (arr[mbinds]
for arr in (obs, returns, masks, actions,
values, neglogpacs))
mb_loss_vals.append(
self._train_step(
lr_now,
cliprange_now,
*slices,
writer=writer,
update=timestep,
cliprange_vf=cliprange_vf_now))
else: # recurrent version
update_fac = max(
self.n_batch // self.nminibatches // self.noptepochs //
self.n_steps, 1)
assert self.n_envs % self.nminibatches == 0
env_indices = np.arange(self.n_envs)
flat_indices = np.arange(self.n_envs *
self.n_steps).reshape(
self.n_envs, self.n_steps)
envs_per_batch = batch_size // self.n_steps
for epoch_num in range(self.noptepochs):
np.random.shuffle(env_indices)
for start in range(0, self.n_envs, envs_per_batch):
timestep = self.num_timesteps // update_fac + (
(epoch_num * self.n_envs + start) //
envs_per_batch)
end = start + envs_per_batch
mb_env_inds = env_indices[start:end]
mb_flat_inds = flat_indices[mb_env_inds].ravel()
slices = (arr[mb_flat_inds]
for arr in (obs, returns, masks, actions,
values, neglogpacs))
mb_states = states[mb_env_inds]
mb_loss_vals.append(
self._train_step(
lr_now,
cliprange_now,
*slices,
update=timestep,
writer=writer,
states=mb_states,
cliprange_vf=cliprange_vf_now))
loss_vals = np.mean(mb_loss_vals, axis=0)
t_now = time.time()
fps = int(self.n_batch / (t_now - t_start))
if writer is not None:
total_episode_reward_logger(
self.episode_reward,
true_reward.reshape((self.n_envs, self.n_steps)),
masks.reshape((self.n_envs, self.n_steps)), writer,
self.num_timesteps)
if self.verbose >= 1 and (update % log_interval == 0
or update == 1):
explained_var = explained_variance(values, returns)
logger.logkv("serial_timesteps", update * self.n_steps)
logger.logkv("n_updates", update)
logger.logkv("total_timesteps", self.num_timesteps)
logger.logkv("fps", fps)
logger.logkv("explained_variance", float(explained_var))
if len(self.ep_info_buf) > 0 and len(
self.ep_info_buf[0]) > 0:
logger.logkv(
'ep_reward_mean',
safe_mean([
ep_info['r'] for ep_info in self.ep_info_buf
]))
logger.logkv(
'ep_len_mean',
safe_mean([
ep_info['l'] for ep_info in self.ep_info_buf
]))
logger.logkv('time_elapsed', t_start - t_first_start)
for (loss_val, loss_name) in zip(loss_vals,
self.loss_names):
logger.logkv(loss_name, loss_val)
logger.dumpkvs()
callback.on_training_end()
return self
def setup_model(self):
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the PPO2 model must be " \
"an instance of common.policies.ActorCriticPolicy."
self.n_batch = self.n_envs * self.n_steps
self.graph = tf.Graph()
with self.graph.as_default():
self.set_random_seed(self.seed)
self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess,
graph=self.graph)
n_batch_step = None
n_batch_train = None
if issubclass(self.policy, RecurrentActorCriticPolicy):
assert self.n_envs % self.nminibatches == 0, "For recurrent policies, "\
"the number of environments run in parallel should be a multiple of nminibatches."
n_batch_step = self.n_envs
n_batch_train = self.n_batch // self.nminibatches
act_model = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
n_batch_step,
reuse=False,
**self.policy_kwargs)
with tf.variable_scope(
"train_model",
reuse=True,
custom_getter=tf_util.outer_scope_getter(
"train_model")):
train_model = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs // self.nminibatches,
self.n_steps,
n_batch_train,
reuse=True,
**self.policy_kwargs)
with tf.variable_scope("loss", reuse=False):
self.action_ph = train_model.pdtype.sample_placeholder(
[None], name="action_ph")
self.advs_ph = tf.placeholder(tf.float32, [None],
name="advs_ph")
self.rewards_ph = tf.placeholder(tf.float32, [None],
name="rewards_ph")
self.old_neglog_pac_ph = tf.placeholder(
tf.float32, [None], name="old_neglog_pac_ph")
self.old_vpred_ph = tf.placeholder(tf.float32, [None],
name="old_vpred_ph")
self.learning_rate_ph = tf.placeholder(
tf.float32, [], name="learning_rate_ph")
self.clip_range_ph = tf.placeholder(tf.float32, [],
name="clip_range_ph")
neglogpac = train_model.proba_distribution.neglogp(
self.action_ph)
self.entropy = tf.reduce_mean(
train_model.proba_distribution.entropy())
vpred = train_model.value_flat
# Value function clipping: not present in the original PPO
if self.cliprange_vf is None:
# Default behavior (legacy from OpenAI baselines):
# use the same clipping as for the policy
self.clip_range_vf_ph = self.clip_range_ph
self.cliprange_vf = self.cliprange
elif isinstance(self.cliprange_vf,
(float, int)) and self.cliprange_vf < 0:
# Original PPO implementation: no value function clipping
self.clip_range_vf_ph = None
else:
# Last possible behavior: clipping range
# specific to the value function
self.clip_range_vf_ph = tf.placeholder(
tf.float32, [], name="clip_range_vf_ph")
if self.clip_range_vf_ph is None:
# No clipping
vpred_clipped = train_model.value_flat
else:
# Clip the different between old and new value
# NOTE: this depends on the reward scaling
vpred_clipped = self.old_vpred_ph + \
tf.clip_by_value(train_model.value_flat - self.old_vpred_ph,
- self.clip_range_vf_ph, self.clip_range_vf_ph)
vf_losses1 = tf.square(vpred - self.rewards_ph)
vf_losses2 = tf.square(vpred_clipped - self.rewards_ph)
self.vf_loss = .5 * tf.reduce_mean(
tf.maximum(vf_losses1, vf_losses2))
ratio = tf.exp(self.old_neglog_pac_ph - neglogpac)
pg_losses = -self.advs_ph * ratio
pg_losses2 = -self.advs_ph * tf.clip_by_value(
ratio, 1.0 - self.clip_range_ph,
1.0 + self.clip_range_ph)
self.pg_loss = tf.reduce_mean(
tf.maximum(pg_losses, pg_losses2))
self.approxkl = .5 * tf.reduce_mean(
tf.square(neglogpac - self.old_neglog_pac_ph))
self.clipfrac = tf.reduce_mean(
tf.cast(
tf.greater(tf.abs(ratio - 1.0),
self.clip_range_ph), tf.float32))
loss = self.pg_loss - self.entropy * self.ent_coef + self.vf_loss * self.vf_coef
tf.summary.scalar('entropy_loss', self.entropy)
tf.summary.scalar('policy_gradient_loss', self.pg_loss)
tf.summary.scalar('value_function_loss', self.vf_loss)
tf.summary.scalar('approximate_kullback-leibler',
self.approxkl)
tf.summary.scalar('clip_factor', self.clipfrac)
tf.summary.scalar('loss', loss)
with tf.variable_scope('model'):
self.params = tf.trainable_variables()
if self.full_tensorboard_log:
for var in self.params:
tf.summary.histogram(var.name, var)
grads = tf.gradients(loss, self.params)
if self.max_grad_norm is not None:
grads, _grad_norm = tf.clip_by_global_norm(
grads, self.max_grad_norm)
grads = list(zip(grads, self.params))
trainer = tf.train.AdamOptimizer(
learning_rate=self.learning_rate_ph, epsilon=1e-5)
self._train = trainer.apply_gradients(grads)
self.loss_names = [
'policy_loss', 'value_loss', 'policy_entropy', 'approxkl',
'clipfrac'
]
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards',
tf.reduce_mean(self.rewards_ph))
tf.summary.scalar('learning_rate',
tf.reduce_mean(self.learning_rate_ph))
tf.summary.scalar('advantage',
tf.reduce_mean(self.advs_ph))
tf.summary.scalar('clip_range',
tf.reduce_mean(self.clip_range_ph))
if self.clip_range_vf_ph is not None:
tf.summary.scalar(
'clip_range_vf',
tf.reduce_mean(self.clip_range_vf_ph))
tf.summary.scalar('old_neglog_action_probability',
tf.reduce_mean(self.old_neglog_pac_ph))
tf.summary.scalar('old_value_pred',
tf.reduce_mean(self.old_vpred_ph))
if self.full_tensorboard_log:
tf.summary.histogram('discounted_rewards',
self.rewards_ph)
tf.summary.histogram('learning_rate',
self.learning_rate_ph)
tf.summary.histogram('advantage', self.advs_ph)
tf.summary.histogram('clip_range', self.clip_range_ph)
tf.summary.histogram('old_neglog_action_probability',
self.old_neglog_pac_ph)
tf.summary.histogram('old_value_pred',
self.old_vpred_ph)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', train_model.obs_ph)
else:
tf.summary.histogram('observation',
train_model.obs_ph)
self.train_model = train_model
self.act_model = act_model
self.step = act_model.step
self.proba_step = act_model.proba_step
self.value = act_model.value
self.initial_state = act_model.initial_state
tf.global_variables_initializer().run(session=self.sess) # pylint: disable=E1101
self.summary = tf.summary.merge_all()
def learn(self, total_timesteps, callback=None, log_interval=100, tb_log_name="ACER",
reset_num_timesteps=True):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
callback = self._init_callback(callback)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn()
self.learning_rate_schedule = Scheduler(initial_value=self.learning_rate, n_values=total_timesteps,
schedule=self.lr_schedule)
episode_stats = EpisodeStats(self.n_steps, self.n_envs)
if self.replay_ratio > 0:
buffer = Buffer(env=self.env, n_steps=self.n_steps, size=self.buffer_size)
else:
buffer = None
t_start = time.time()
callback.on_training_start(locals(), globals())
# n_batch samples, 1 on_policy call and multiple off-policy calls
for steps in range(0, total_timesteps, self.n_batch):
callback.on_rollout_start()
enc_obs, obs, actions, rewards, mus, dones, masks = self.runner.run(callback)
callback.update_locals(locals())
callback.on_rollout_end()
# Early stopping due to the callback
if not self.runner.continue_training:
break
episode_stats.feed(rewards, dones)
if buffer is not None:
buffer.put(enc_obs, actions, rewards, mus, dones, masks)
if writer is not None:
total_episode_reward_logger(self.episode_reward,
rewards.reshape((self.n_envs, self.n_steps)),
dones.reshape((self.n_envs, self.n_steps)),
writer, self.num_timesteps)
# reshape stuff correctly
obs = obs.reshape(self.runner.batch_ob_shape)
actions = actions.reshape([self.n_batch])
rewards = rewards.reshape([self.n_batch])
mus = mus.reshape([self.n_batch, self.n_act])
dones = dones.reshape([self.n_batch])
masks = masks.reshape([self.runner.batch_ob_shape[0]])
names_ops, values_ops = self._train_step(obs, actions, rewards, dones, mus, self.initial_state, masks,
self.num_timesteps, writer)
if self.verbose >= 1 and (int(steps / self.n_batch) % log_interval == 0):
logger.record_tabular("total_timesteps", self.num_timesteps)
logger.record_tabular("fps", int(steps / (time.time() - t_start)))
# IMP: In EpisodicLife env, during training, we get done=True at each loss of life,
# not just at the terminal state. Thus, this is mean until end of life, not end of episode.
# For true episode rewards, see the monitor files in the log folder.
logger.record_tabular("mean_episode_length", episode_stats.mean_length())
logger.record_tabular("mean_episode_reward", episode_stats.mean_reward())
for name, val in zip(names_ops, values_ops):
logger.record_tabular(name, float(val))
logger.dump_tabular()
if (self.replay_ratio > 0 and
buffer is not None and
buffer.has_atleast(self.replay_start)):
samples_number = np.random.poisson(self.replay_ratio)
for _ in range(samples_number):
# get obs, actions, rewards, mus, dones from buffer.
obs, actions, rewards, mus, dones, masks = buffer.get()
# reshape stuff correctly
obs = obs.reshape(self.runner.batch_ob_shape)
actions = actions.reshape([self.n_batch])
rewards = rewards.reshape([self.n_batch])
mus = mus.reshape([self.n_batch, self.n_act])
dones = dones.reshape([self.n_batch])
masks = masks.reshape([self.runner.batch_ob_shape[0]])
self._train_step(obs, actions, rewards, dones, mus, self.initial_state, masks,
self.num_timesteps)
callback.on_training_end()
return self
def setup_model(self):
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the ACER model must be " \
"an instance of common.policies.ActorCriticPolicy."
if isinstance(self.action_space, Discrete):
self.n_act = self.action_space.n
continuous = False
elif isinstance(self.action_space, Box):
# self.n_act = self.action_space.shape[-1]
# continuous = True
raise NotImplementedError("WIP: Acer does not support Continuous actions yet.")
else:
raise ValueError("Error: ACER does not work with {} actions space.".format(self.action_space))
self.n_batch = self.n_envs * self.n_steps
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess, graph=self.graph)
self.set_random_seed(self.seed)
n_batch_step = None
if issubclass(self.policy, RecurrentActorCriticPolicy):
n_batch_step = self.n_envs
n_batch_train = self.n_envs * (self.n_steps + 1)
step_model = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs, 1,
n_batch_step, reuse=False, **self.policy_kwargs)
self.params = tf_util.get_trainable_vars("model")
with tf.variable_scope("train_model", reuse=True,
custom_getter=tf_util.outer_scope_getter("train_model")):
train_model = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs,
self.n_steps + 1, n_batch_train, reuse=True, **self.policy_kwargs)
with tf.variable_scope("moving_average"):
# create averaged model
ema = tf.train.ExponentialMovingAverage(self.alpha)
ema_apply_op = ema.apply(self.params)
def custom_getter(getter, name, *args, **kwargs):
name = name.replace("polyak_model/", "")
val = ema.average(getter(name, *args, **kwargs))
return val
with tf.variable_scope("polyak_model", reuse=True, custom_getter=custom_getter):
self.polyak_model = polyak_model = self.policy(self.sess, self.observation_space, self.action_space,
self.n_envs, self.n_steps + 1,
self.n_envs * (self.n_steps + 1), reuse=True,
**self.policy_kwargs)
with tf.variable_scope("loss", reuse=False):
self.done_ph = tf.placeholder(tf.float32, [self.n_batch]) # dones
self.reward_ph = tf.placeholder(tf.float32, [self.n_batch]) # rewards, not returns
self.mu_ph = tf.placeholder(tf.float32, [self.n_batch, self.n_act]) # mu's
self.action_ph = train_model.pdtype.sample_placeholder([self.n_batch])
self.learning_rate_ph = tf.placeholder(tf.float32, [])
eps = 1e-6
# Notation: (var) = batch variable, (var)s = sequence variable,
# (var)_i = variable index by action at step i
# shape is [n_envs * (n_steps + 1)]
if continuous:
value = train_model.value_flat
else:
value = tf.reduce_sum(train_model.policy_proba * train_model.q_value, axis=-1)
rho, rho_i_ = None, None
if continuous:
action_ = strip(train_model.proba_distribution.sample(), self.n_envs, self.n_steps)
distribution_f = tf.contrib.distributions.MultivariateNormalDiag(
loc=strip(train_model.proba_distribution.mean, self.n_envs, self.n_steps),
scale_diag=strip(train_model.proba_distribution.logstd, self.n_envs, self.n_steps))
f_polyak = tf.contrib.distributions.MultivariateNormalDiag(
loc=strip(polyak_model.proba_distribution.mean, self.n_envs, self.n_steps),
scale_diag=strip(polyak_model.proba_distribution.logstd, self.n_envs, self.n_steps))
f_i = distribution_f.prob(self.action_ph)
f_i_ = distribution_f.prob(action_)
f_polyak_i = f_polyak.prob(self.action_ph)
phi_i = strip(train_model.proba_distribution.mean, self.n_envs, self.n_steps)
q_value = strip(train_model.value_fn, self.n_envs, self.n_steps)
q_i = q_value[:, 0]
rho_i = tf.reshape(f_i, [-1, 1]) / (self.mu_ph + eps)
rho_i_ = tf.reshape(f_i_, [-1, 1]) / (self.mu_ph + eps)
qret = q_retrace(self.reward_ph, self.done_ph, q_i, value, tf.pow(rho_i, 1 / self.n_act),
self.n_envs, self.n_steps, self.gamma)
else:
# strip off last step
# f is a distribution, chosen to be Gaussian distributions
# with fixed diagonal covariance and mean \phi(x)
# in the paper
distribution_f, f_polyak, q_value = \
map(lambda variables: strip(variables, self.n_envs, self.n_steps),
[train_model.policy_proba, polyak_model.policy_proba, train_model.q_value])
# Get pi and q values for actions taken
f_i = get_by_index(distribution_f, self.action_ph)
f_i_ = distribution_f
phi_i = distribution_f
f_polyak_i = f_polyak
q_i = get_by_index(q_value, self.action_ph)
# Compute ratios for importance truncation
rho = distribution_f / (self.mu_ph + eps)
rho_i = get_by_index(rho, self.action_ph)
# Calculate Q_retrace targets
qret = q_retrace(self.reward_ph, self.done_ph, q_i, value, rho_i, self.n_envs, self.n_steps,
self.gamma)
# Calculate losses
# Entropy
entropy = tf.reduce_sum(train_model.proba_distribution.entropy())
# Policy Gradient loss, with truncated importance sampling & bias correction
value = strip(value, self.n_envs, self.n_steps, True)
# check_shape([qret, value, rho_i, f_i], [[self.n_envs * self.n_steps]] * 4)
# check_shape([rho, distribution_f, q_value], [[self.n_envs * self.n_steps, self.n_act]] * 2)
# Truncated importance sampling
adv = qret - value
log_f = tf.log(f_i + eps)
# [n_envs * n_steps]
gain_f = log_f * tf.stop_gradient(adv * tf.minimum(self.correction_term, rho_i))
loss_f = -tf.reduce_mean(gain_f)
# Bias correction for the truncation
adv_bc = (q_value - tf.reshape(value, [self.n_envs * self.n_steps, 1])) # [n_envs * n_steps, n_act]
# check_shape([adv_bc, log_f_bc], [[self.n_envs * self.n_steps, self.n_act]] * 2)
if continuous:
gain_bc = tf.stop_gradient(adv_bc *
tf.nn.relu(1.0 - (self.correction_term / (rho_i_ + eps))) *
f_i_)
else:
log_f_bc = tf.log(f_i_ + eps) # / (f_old + eps)
gain_bc = tf.reduce_sum(log_f_bc *
tf.stop_gradient(
adv_bc *
tf.nn.relu(1.0 - (self.correction_term / (rho + eps))) *
f_i_),
axis=1)
# IMP: This is sum, as expectation wrt f
loss_bc = -tf.reduce_mean(gain_bc)
loss_policy = loss_f + loss_bc
# Value/Q function loss, and explained variance
check_shape([qret, q_i], [[self.n_envs * self.n_steps]] * 2)
explained_variance = q_explained_variance(tf.reshape(q_i, [self.n_envs, self.n_steps]),
tf.reshape(qret, [self.n_envs, self.n_steps]))
loss_q = tf.reduce_mean(tf.square(tf.stop_gradient(qret) - q_i) * 0.5)
# Net loss
check_shape([loss_policy, loss_q, entropy], [[]] * 3)
loss = loss_policy + self.q_coef * loss_q - self.ent_coef * entropy
tf.summary.scalar('entropy_loss', entropy)
tf.summary.scalar('policy_gradient_loss', loss_policy)
tf.summary.scalar('value_function_loss', loss_q)
tf.summary.scalar('loss', loss)
norm_grads_q, norm_grads_policy, avg_norm_grads_f = None, None, None
avg_norm_k, avg_norm_g, avg_norm_k_dot_g, avg_norm_adj = None, None, None, None
if self.trust_region:
# [n_envs * n_steps, n_act]
grad = tf.gradients(- (loss_policy - self.ent_coef * entropy) * self.n_steps * self.n_envs,
phi_i)
# [n_envs * n_steps, n_act] # Directly computed gradient of KL divergence wrt f
kl_grad = - f_polyak_i / (f_i_ + eps)
k_dot_g = tf.reduce_sum(kl_grad * grad, axis=-1)
adj = tf.maximum(0.0, (tf.reduce_sum(kl_grad * grad, axis=-1) - self.delta) / (
tf.reduce_sum(tf.square(kl_grad), axis=-1) + eps)) # [n_envs * n_steps]
# Calculate stats (before doing adjustment) for logging.
avg_norm_k = avg_norm(kl_grad)
avg_norm_g = avg_norm(grad)
avg_norm_k_dot_g = tf.reduce_mean(tf.abs(k_dot_g))
avg_norm_adj = tf.reduce_mean(tf.abs(adj))
grad = grad - tf.reshape(adj, [self.n_envs * self.n_steps, 1]) * kl_grad
# These are turst region adjusted gradients wrt f ie statistics of policy pi
grads_f = -grad / (self.n_envs * self.n_steps)
grads_policy = tf.gradients(f_i_, self.params, grads_f)
grads_q = tf.gradients(loss_q * self.q_coef, self.params)
grads = [gradient_add(g1, g2, param, verbose=self.verbose)
for (g1, g2, param) in zip(grads_policy, grads_q, self.params)]
avg_norm_grads_f = avg_norm(grads_f) * (self.n_steps * self.n_envs)
norm_grads_q = tf.global_norm(grads_q)
norm_grads_policy = tf.global_norm(grads_policy)
else:
grads = tf.gradients(loss, self.params)
norm_grads = None
if self.max_grad_norm is not None:
grads, norm_grads = tf.clip_by_global_norm(grads, self.max_grad_norm)
grads = list(zip(grads, self.params))
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('rewards', tf.reduce_mean(self.reward_ph))
tf.summary.scalar('learning_rate', tf.reduce_mean(self.learning_rate))
tf.summary.scalar('advantage', tf.reduce_mean(adv))
tf.summary.scalar('action_probability', tf.reduce_mean(self.mu_ph))
if self.full_tensorboard_log:
tf.summary.histogram('rewards', self.reward_ph)
tf.summary.histogram('learning_rate', self.learning_rate)
tf.summary.histogram('advantage', adv)
tf.summary.histogram('action_probability', self.mu_ph)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', train_model.obs_ph)
else:
tf.summary.histogram('observation', train_model.obs_ph)
trainer = tf.train.RMSPropOptimizer(learning_rate=self.learning_rate_ph, decay=self.rprop_alpha,
epsilon=self.rprop_epsilon)
_opt_op = trainer.apply_gradients(grads)
# so when you call _train, you first do the gradient step, then you apply ema
with tf.control_dependencies([_opt_op]):
_train = tf.group(ema_apply_op)
# Ops/Summaries to run, and their names for logging
assert norm_grads is not None
run_ops = [_train, loss, loss_q, entropy, loss_policy, loss_f, loss_bc, explained_variance, norm_grads]
names_ops = ['loss', 'loss_q', 'entropy', 'loss_policy', 'loss_f', 'loss_bc', 'explained_variance',
'norm_grads']
if self.trust_region:
self.run_ops = run_ops + [norm_grads_q, norm_grads_policy, avg_norm_grads_f, avg_norm_k, avg_norm_g,
avg_norm_k_dot_g, avg_norm_adj]
self.names_ops = names_ops + ['norm_grads_q', 'norm_grads_policy', 'avg_norm_grads_f', 'avg_norm_k',
'avg_norm_g', 'avg_norm_k_dot_g', 'avg_norm_adj']
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.proba_step = step_model.proba_step
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=self.sess)
self.summary = tf.summary.merge_all()
