def _create_mu_log_sigma(
self,
logits: tf.Tensor,
act_size: List[int],
log_sigma_min: float,
log_sigma_max: float,
) -> "GaussianDistribution.MuSigmaTensors":
mu = tf.layers.dense(
logits,
act_size[0],
activation=None,
name="mu",
kernel_initializer=ModelUtils.scaled_init(0.01),
reuse=tf.AUTO_REUSE,
)
# Policy-dependent log_sigma_sq
log_sigma = tf.layers.dense(
logits,
act_size[0],
activation=None,
name="log_std",
kernel_initializer=ModelUtils.scaled_init(0.01),
)
log_sigma = tf.clip_by_value(log_sigma, log_sigma_min, log_sigma_max)
sigma = tf.exp(log_sigma)
return self.MuSigmaTensors(mu, log_sigma, sigma)
def create_loss(self, learning_rate: float) -> None:
"""
Creates the loss and update nodes for the GAIL reward generator
:param learning_rate: The learning rate for the optimizer
"""
self.mean_expert_estimate = tf.reduce_mean(self.expert_estimate)
self.mean_policy_estimate = tf.reduce_mean(self.policy_estimate)
if self.use_vail:
self.beta = tf.get_variable(
"gail_beta",
[],
trainable=False,
dtype=tf.float32,
initializer=tf.ones_initializer(),
)
self.discriminator_loss = -tf.reduce_mean(
tf.log(self.expert_estimate + EPSILON)
+ tf.log(1.0 - self.policy_estimate + EPSILON)
)
if self.use_vail:
# KL divergence loss (encourage latent representation to be normal)
self.kl_loss = tf.reduce_mean(
-tf.reduce_sum(
1
+ self.z_log_sigma_sq
- 0.5 * tf.square(self.z_mean_expert)
- 0.5 * tf.square(self.z_mean_policy)
- tf.exp(self.z_log_sigma_sq),
1,
)
)
self.loss = (
self.beta * (self.kl_loss - self.mutual_information)
+ self.discriminator_loss
)
else:
self.loss = self.discriminator_loss
if self.gradient_penalty_weight > 0.0:
self.loss += self.gradient_penalty_weight * self.create_gradient_magnitude()
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
self.update_batch = optimizer.minimize(self.loss)
def _create_losses(self, probs, old_probs, value_heads, entropy, beta,
epsilon, lr, max_step):
"""
Creates training-specific Tensorflow ops for PPO models.
:param probs: Current policy probabilities
:param old_probs: Past policy probabilities
:param value_heads: Value estimate tensors from each value stream
:param beta: Entropy regularization strength
:param entropy: Current policy entropy
:param epsilon: Value for policy-divergence threshold
:param lr: Learning rate
:param max_step: Total number of training steps.
"""
self.returns_holders = {}
self.old_values = {}
for name in value_heads.keys():
returns_holder = tf.placeholder(shape=[None],
dtype=tf.float32,
name="{}_returns".format(name))
old_value = tf.placeholder(shape=[None],
dtype=tf.float32,
name="{}_value_estimate".format(name))
self.returns_holders[name] = returns_holder
self.old_values[name] = old_value
self.advantage = tf.placeholder(shape=[None],
dtype=tf.float32,
name="advantages")
advantage = tf.expand_dims(self.advantage, -1)
decay_epsilon = tf.train.polynomial_decay(epsilon,
self.policy.global_step,
max_step,
0.1,
power=1.0)
decay_beta = tf.train.polynomial_decay(beta,
self.policy.global_step,
max_step,
1e-5,
power=1.0)
value_losses = []
for name, head in value_heads.items():
clipped_value_estimate = self.old_values[name] + tf.clip_by_value(
tf.reduce_sum(head, axis=1) - self.old_values[name],
-decay_epsilon,
decay_epsilon,
)
v_opt_a = tf.squared_difference(self.returns_holders[name],
tf.reduce_sum(head, axis=1))
v_opt_b = tf.squared_difference(self.returns_holders[name],
clipped_value_estimate)
value_loss = tf.reduce_mean(
tf.dynamic_partition(tf.maximum(v_opt_a, v_opt_b),
self.policy.mask, 2)[1])
value_losses.append(value_loss)
self.value_loss = tf.reduce_mean(value_losses)
r_theta = tf.exp(probs - old_probs)
p_opt_a = r_theta * advantage
p_opt_b = (tf.clip_by_value(r_theta, 1.0 - decay_epsilon,
1.0 + decay_epsilon) * advantage)
self.policy_loss = -tf.reduce_mean(
tf.dynamic_partition(tf.minimum(p_opt_a, p_opt_b),
self.policy.mask, 2)[1])
# For cleaner stats reporting
self.abs_policy_loss = tf.abs(self.policy_loss)
self.loss = (
self.policy_loss + 0.5 * self.value_loss -
decay_beta * tf.reduce_mean(
tf.dynamic_partition(entropy, self.policy.mask, 2)[1]))
def _create_losses(
self,
q1_streams: Dict[str, tf.Tensor],
q2_streams: Dict[str, tf.Tensor],
lr: tf.Tensor,
max_step: int,
stream_names: List[str],
discrete: bool = False,
) -> None:
"""
Creates training-specific Tensorflow ops for SAC models.
:param q1_streams: Q1 streams from policy network
:param q1_streams: Q2 streams from policy network
:param lr: Learning rate
:param max_step: Total number of training steps.
:param stream_names: List of reward stream names.
:param discrete: Whether or not to use discrete action losses.
"""
if discrete:
self.target_entropy = [
self.discrete_target_entropy_scale *
np.log(i).astype(np.float32) for i in self.act_size
]
discrete_action_probs = tf.exp(self.policy.all_log_probs)
per_action_entropy = discrete_action_probs * self.policy.all_log_probs
else:
self.target_entropy = (
-1 * self.continuous_target_entropy_scale *
np.prod(self.act_size[0]).astype(np.float32))
self.rewards_holders = {}
self.min_policy_qs = {}
for name in stream_names:
if discrete:
_branched_mpq1 = ModelUtils.break_into_branches(
self.policy_network.q1_pheads[name] *
discrete_action_probs,
self.act_size,
)
branched_mpq1 = tf.stack([
tf.reduce_sum(_br, axis=1, keep_dims=True)
for _br in _branched_mpq1
])
_q1_p_mean = tf.reduce_mean(branched_mpq1, axis=0)
_branched_mpq2 = ModelUtils.break_into_branches(
self.policy_network.q2_pheads[name] *
discrete_action_probs,
self.act_size,
)
branched_mpq2 = tf.stack([
tf.reduce_sum(_br, axis=1, keep_dims=True)
for _br in _branched_mpq2
])
_q2_p_mean = tf.reduce_mean(branched_mpq2, axis=0)
self.min_policy_qs[name] = tf.minimum(_q1_p_mean, _q2_p_mean)
else:
self.min_policy_qs[name] = tf.minimum(
self.policy_network.q1_pheads[name],
self.policy_network.q2_pheads[name],
)
rewards_holder = tf.placeholder(shape=[None],
dtype=tf.float32,
name=f"{name}_rewards")
self.rewards_holders[name] = rewards_holder
q1_losses = []
q2_losses = []
# Multiple q losses per stream
expanded_dones = tf.expand_dims(self.dones_holder, axis=-1)
for i, name in enumerate(stream_names):
_expanded_rewards = tf.expand_dims(self.rewards_holders[name],
axis=-1)
q_backup = tf.stop_gradient(
_expanded_rewards +
(1.0 - self.use_dones_in_backup[name] * expanded_dones) *
self.gammas[i] * self.target_network.value_heads[name])
if discrete:
# We need to break up the Q functions by branch, and update them individually.
branched_q1_stream = ModelUtils.break_into_branches(
self.policy.selected_actions * q1_streams[name],
self.act_size)
branched_q2_stream = ModelUtils.break_into_branches(
self.policy.selected_actions * q2_streams[name],
self.act_size)
# Reduce each branch into scalar
branched_q1_stream = [
tf.reduce_sum(_branch, axis=1, keep_dims=True)
for _branch in branched_q1_stream
]
branched_q2_stream = [
tf.reduce_sum(_branch, axis=1, keep_dims=True)
for _branch in branched_q2_stream
]
q1_stream = tf.reduce_mean(branched_q1_stream, axis=0)
q2_stream = tf.reduce_mean(branched_q2_stream, axis=0)
else:
q1_stream = q1_streams[name]
q2_stream = q2_streams[name]
_q1_loss = 0.5 * tf.reduce_mean(
tf.to_float(self.policy.mask) *
tf.squared_difference(q_backup, q1_stream))
_q2_loss = 0.5 * tf.reduce_mean(
tf.to_float(self.policy.mask) *
tf.squared_difference(q_backup, q2_stream))
q1_losses.append(_q1_loss)
q2_losses.append(_q2_loss)
self.q1_loss = tf.reduce_mean(q1_losses)
self.q2_loss = tf.reduce_mean(q2_losses)
# Learn entropy coefficient
if discrete:
# Create a log_ent_coef for each branch
self.log_ent_coef = tf.get_variable(
"log_ent_coef",
dtype=tf.float32,
initializer=np.log([self.init_entcoef] *
len(self.act_size)).astype(np.float32),
trainable=True,
)
else:
self.log_ent_coef = tf.get_variable(
"log_ent_coef",
dtype=tf.float32,
initializer=np.log(self.init_entcoef).astype(np.float32),
trainable=True,
)
self.ent_coef = tf.exp(self.log_ent_coef)
if discrete:
# We also have to do a different entropy and target_entropy per branch.
branched_per_action_ent = ModelUtils.break_into_branches(
per_action_entropy, self.act_size)
branched_ent_sums = tf.stack(
[
tf.reduce_sum(_lp, axis=1, keep_dims=True) + _te for _lp,
_te in zip(branched_per_action_ent, self.target_entropy)
],
axis=1,
)
self.entropy_loss = -tf.reduce_mean(
tf.to_float(self.policy.mask) * tf.reduce_mean(
self.log_ent_coef *
tf.squeeze(tf.stop_gradient(branched_ent_sums), axis=2),
axis=1,
))
# Same with policy loss, we have to do the loss per branch and average them,
# so that larger branches don't get more weight.
# The equivalent KL divergence from Eq 10 of Haarnoja et al. is also pi*log(pi) - Q
branched_q_term = ModelUtils.break_into_branches(
discrete_action_probs * self.policy_network.q1_p,
self.act_size)
branched_policy_loss = tf.stack([
tf.reduce_sum(self.ent_coef[i] * _lp - _qt,
axis=1,
keep_dims=True)
for i, (_lp, _qt) in enumerate(
zip(branched_per_action_ent, branched_q_term))
])
self.policy_loss = tf.reduce_mean(
tf.to_float(self.policy.mask) *
tf.squeeze(branched_policy_loss))
# Do vbackup entropy bonus per branch as well.
branched_ent_bonus = tf.stack([
tf.reduce_sum(self.ent_coef[i] * _lp, axis=1, keep_dims=True)
for i, _lp in enumerate(branched_per_action_ent)
])
value_losses = []
for name in stream_names:
v_backup = tf.stop_gradient(
self.min_policy_qs[name] -
tf.reduce_mean(branched_ent_bonus, axis=0))
value_losses.append(0.5 * tf.reduce_mean(
tf.to_float(self.policy.mask) * tf.squared_difference(
self.policy_network.value_heads[name], v_backup)))
else:
self.entropy_loss = -tf.reduce_mean(
self.log_ent_coef * tf.to_float(self.policy.mask) *
tf.stop_gradient(
tf.reduce_sum(
self.policy.all_log_probs + self.target_entropy,
axis=1,
keep_dims=True,
)))
batch_policy_loss = tf.reduce_mean(
self.ent_coef * self.policy.all_log_probs -
self.policy_network.q1_p,
axis=1,
)
self.policy_loss = tf.reduce_mean(
tf.to_float(self.policy.mask) * batch_policy_loss)
value_losses = []
for name in stream_names:
v_backup = tf.stop_gradient(
self.min_policy_qs[name] - tf.reduce_sum(
self.ent_coef * self.policy.all_log_probs, axis=1))
value_losses.append(0.5 * tf.reduce_mean(
tf.to_float(self.policy.mask) * tf.squared_difference(
self.policy_network.value_heads[name], v_backup)))
self.value_loss = tf.reduce_mean(value_losses)
self.total_value_loss = self.q1_loss + self.q2_loss + self.value_loss
self.entropy = self.policy_network.entropy
def create_cc_actor(self, hidden_policy, scope):
"""
Creates Continuous control actor for SAC.
:param hidden_policy: Output of feature extractor (i.e. the input for vector obs, output of CNN for visual obs).
:param num_layers: TF scope to assign whatever is created in this block.
"""
# Create action input (continuous)
self.action_holder = tf.placeholder(shape=[None, self.act_size[0]],
dtype=tf.float32,
name="action_holder")
self.external_action_in = self.action_holder
scope = self.join_scopes(scope, "policy")
with tf.variable_scope(scope):
hidden_policy = self.create_vector_observation_encoder(
hidden_policy,
self.h_size,
self.activ_fn,
self.num_layers,
"encoder",
False,
)
if self.use_recurrent:
hidden_policy, memory_out = self.create_recurrent_encoder(
hidden_policy,
self.policy_memory_in,
self.sequence_length,
name="lstm_policy",
)
self.policy_memory_out = memory_out
with tf.variable_scope(scope):
mu = tf.layers.dense(
hidden_policy,
self.act_size[0],
activation=None,
name="mu",
kernel_initializer=LearningModel.scaled_init(0.01),
)
# Policy-dependent log_sigma_sq
log_sigma_sq = tf.layers.dense(
hidden_policy,
self.act_size[0],
activation=None,
name="log_std",
kernel_initializer=LearningModel.scaled_init(0.01),
)
self.log_sigma_sq = tf.clip_by_value(log_sigma_sq, LOG_STD_MIN,
LOG_STD_MAX)
sigma_sq = tf.exp(self.log_sigma_sq)
# Do the reparameterization trick
policy_ = mu + tf.random_normal(tf.shape(mu)) * sigma_sq
_gauss_pre = -0.5 * (((policy_ - mu) /
(tf.exp(self.log_sigma_sq) + EPSILON))**2 +
2 * self.log_sigma_sq + np.log(2 * np.pi))
all_probs = tf.reduce_sum(_gauss_pre, axis=1, keepdims=True)
self.entropy = tf.reduce_sum(self.log_sigma_sq +
0.5 * np.log(2.0 * np.pi * np.e),
axis=-1)
# Squash probabilities
# Keep deterministic around in case we want to use it.
self.deterministic_output = tf.tanh(mu)
# Note that this is just for symmetry with PPO.
self.output_pre = tf.tanh(policy_)
# Squash correction
all_probs -= tf.reduce_sum(tf.log(1 - self.output_pre**2 +
EPSILON),
axis=1,
keepdims=True)
self.all_log_probs = all_probs
self.selected_actions = tf.stop_gradient(self.output_pre)
self.action_probs = all_probs
# Extract output for Barracuda
self.output = tf.identity(self.output_pre, name="action")
# Get all policy vars
self.policy_vars = self.get_vars(scope)
def create_cc_actor_critic(self, h_size: int, num_layers: int,
vis_encode_type: EncoderType) -> None:
"""
Creates Continuous control actor-critic model.
:param h_size: Size of hidden linear layers.
:param num_layers: Number of hidden linear layers.
"""
hidden_streams = self.create_observation_streams(
2, h_size, num_layers, vis_encode_type)
if self.use_recurrent:
self.memory_in = tf.placeholder(shape=[None, self.m_size],
dtype=tf.float32,
name="recurrent_in")
_half_point = int(self.m_size / 2)
hidden_policy, memory_policy_out = self.create_recurrent_encoder(
hidden_streams[0],
self.memory_in[:, :_half_point],
self.sequence_length,
name="lstm_policy",
)
hidden_value, memory_value_out = self.create_recurrent_encoder(
hidden_streams[1],
self.memory_in[:, _half_point:],
self.sequence_length,
name="lstm_value",
)
self.memory_out = tf.concat([memory_policy_out, memory_value_out],
axis=1,
name="recurrent_out")
else:
hidden_policy = hidden_streams[0]
hidden_value = hidden_streams[1]
mu = tf.layers.dense(
hidden_policy,
self.act_size[0],
activation=None,
kernel_initializer=LearningModel.scaled_init(0.01),
reuse=tf.AUTO_REUSE,
)
self.log_sigma_sq = tf.get_variable(
"log_sigma_squared",
[self.act_size[0]],
dtype=tf.float32,
initializer=tf.zeros_initializer(),
)
sigma_sq = tf.exp(self.log_sigma_sq)
self.epsilon = tf.placeholder(shape=[None, self.act_size[0]],
dtype=tf.float32,
name="epsilon")
# Clip and scale output to ensure actions are always within [-1, 1] range.
self.output_pre = mu + tf.sqrt(sigma_sq) * self.epsilon
output_post = tf.clip_by_value(self.output_pre, -3, 3) / 3
self.output = tf.identity(output_post, name="action")
self.selected_actions = tf.stop_gradient(output_post)
# Compute probability of model output.
all_probs = (-0.5 * tf.square(tf.stop_gradient(self.output_pre) - mu) /
sigma_sq - 0.5 * tf.log(2.0 * np.pi) -
0.5 * self.log_sigma_sq)
self.all_log_probs = tf.identity(all_probs, name="action_probs")
self.entropy = 0.5 * tf.reduce_mean(
tf.log(2 * np.pi * np.e) + self.log_sigma_sq)
self.create_value_heads(self.stream_names, hidden_value)
self.all_old_log_probs = tf.placeholder(shape=[None, self.act_size[0]],
dtype=tf.float32,
name="old_probabilities")
# We keep these tensors the same name, but use new nodes to keep code parallelism with discrete control.
self.log_probs = tf.reduce_sum((tf.identity(self.all_log_probs)),
axis=1,
keepdims=True)
self.old_log_probs = tf.reduce_sum(
(tf.identity(self.all_old_log_probs)), axis=1, keepdims=True)
def _create_cc_actor(
self,
encoded: tf.Tensor,
tanh_squash: bool = False,
reparameterize: bool = False,
condition_sigma_on_obs: bool = True,
) -> None:
"""
Creates Continuous control actor-critic model.
:param h_size: Size of hidden linear layers.
:param num_layers: Number of hidden linear layers.
:param vis_encode_type: Type of visual encoder to use if visual input.
:param tanh_squash: Whether to use a tanh function, or a clipped output.
:param reparameterize: Whether we are using the resampling trick to update the policy.
"""
if self.use_recurrent:
self.memory_in = tf.placeholder(shape=[None, self.m_size],
dtype=tf.float32,
name="recurrent_in")
hidden_policy, memory_policy_out = ModelUtils.create_recurrent_encoder(
encoded,
self.memory_in,
self.sequence_length_ph,
name="lstm_policy")
self.memory_out = tf.identity(memory_policy_out,
name="recurrent_out")
else:
hidden_policy = encoded
with tf.variable_scope("policy"):
mu = tf.layers.dense(
hidden_policy,
self.act_size[0],
activation=None,
name="mu",
kernel_initializer=ModelUtils.scaled_init(0.01),
reuse=tf.AUTO_REUSE,
)
# Policy-dependent log_sigma
if condition_sigma_on_obs:
log_sigma = tf.layers.dense(
hidden_policy,
self.act_size[0],
activation=None,
name="log_sigma",
kernel_initializer=ModelUtils.scaled_init(0.01),
)
else:
log_sigma = tf.get_variable(
"log_sigma",
[self.act_size[0]],
dtype=tf.float32,
initializer=tf.zeros_initializer(),
)
log_sigma = tf.clip_by_value(log_sigma, self.log_std_min,
self.log_std_max)
sigma = tf.exp(log_sigma)
epsilon = tf.random_normal(tf.shape(mu))
sampled_policy = mu + sigma * epsilon
# Stop gradient if we're not doing the resampling trick
if not reparameterize:
sampled_policy_probs = tf.stop_gradient(sampled_policy)
else:
sampled_policy_probs = sampled_policy
# Compute probability of model output.
_gauss_pre = -0.5 * (
((sampled_policy_probs - mu) /
(sigma + EPSILON))**2 + 2 * log_sigma + np.log(2 * np.pi))
all_probs = _gauss_pre
all_probs = tf.reduce_sum(_gauss_pre, axis=1, keepdims=True)
if tanh_squash:
self.output_pre = tf.tanh(sampled_policy)
# Squash correction
all_probs -= tf.reduce_sum(tf.log(1 - self.output_pre**2 +
EPSILON),
axis=1,
keepdims=True)
self.output = tf.identity(self.output_pre, name="action")
else:
self.output_pre = sampled_policy
# Clip and scale output to ensure actions are always within [-1, 1] range.
output_post = tf.clip_by_value(self.output_pre, -3, 3) / 3
self.output = tf.identity(output_post, name="action")
self.selected_actions = tf.stop_gradient(self.output)
self.all_log_probs = tf.identity(all_probs, name="action_probs")
single_dim_entropy = 0.5 * tf.reduce_mean(
tf.log(2 * np.pi * np.e) + 2 * log_sigma)
# Make entropy the right shape
self.entropy = tf.ones_like(tf.reshape(mu[:, 0],
[-1])) * single_dim_entropy
# We keep these tensors the same name, but use new nodes to keep code parallelism with discrete control.
self.log_probs = tf.reduce_sum((tf.identity(self.all_log_probs)),
axis=1,
keepdims=True)
self.action_holder = tf.placeholder(shape=[None, self.act_size[0]],
dtype=tf.float32,
name="action_holder")
