def q_retrace(rewards, dones, q_i, values, rho_i, n_envs, n_steps, gamma):
"""
Calculates the target Q-retrace
:param rewards: ([TensorFlow Tensor]) The rewards
:param dones: ([TensorFlow Tensor])
:param q_i: ([TensorFlow Tensor]) The Q values for actions taken
:param values: ([TensorFlow Tensor]) The output of the value functions
:param rho_i: ([TensorFlow Tensor]) The importance weight for each action
:param n_envs: (int) The number of environments
:param n_steps: (int) The number of steps to run for each environment
:param gamma: (float) The discount value
:return: ([TensorFlow Tensor]) the target Q-retrace
"""
rho_bar = batch_to_seq(tf.minimum(1.0, rho_i), n_envs, n_steps, True) # list of len steps, shape [n_envs]
reward_seq = batch_to_seq(rewards, n_envs, n_steps, True) # list of len steps, shape [n_envs]
done_seq = batch_to_seq(dones, n_envs, n_steps, True) # list of len steps, shape [n_envs]
q_is = batch_to_seq(q_i, n_envs, n_steps, True)
value_sequence = batch_to_seq(values, n_envs, n_steps + 1, True)
final_value = value_sequence[-1]
qret = final_value
qrets = []
for i in range(n_steps - 1, -1, -1):
check_shape([qret, done_seq[i], reward_seq[i], rho_bar[i], q_is[i], value_sequence[i]], [[n_envs]] * 6)
qret = reward_seq[i] + gamma * qret * (1.0 - done_seq[i])
qrets.append(qret)
qret = (rho_bar[i] * (qret - q_is[i])) + value_sequence[i]
qrets = qrets[::-1]
qret = seq_to_batch(qrets, flat=True)
return qret
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.num_procs,
graph=self.graph)
n_batch_step = None
if issubclass(self.policy, LstmPolicy):
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.params = find_trainable_variables("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)
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)
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_fn[:, 0]
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.histogram('rewards', self.reward_ph)
tf.summary.scalar('learning_rate',
tf.reduce_mean(self.learning_rate))
tf.summary.histogram('learning_rate', self.learning_rate)
tf.summary.scalar('advantage', tf.reduce_mean(adv))
tf.summary.histogram('advantage', adv)
tf.summary.scalar('action_probabilty',
tf.reduce_mean(self.mu_ph))
tf.summary.histogram('action_probabilty', self.mu_ph)
if len(self.observation_space.shape) == 3:
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()
