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_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"):
distribution = GaussianDistribution(
hidden_policy,
self.act_size,
reparameterize=reparameterize,
tanh_squash=tanh_squash,
condition_sigma=condition_sigma_on_obs,
)
if tanh_squash:
self.output_pre = distribution.sample
self.output = tf.identity(self.output_pre, name="action")
else:
self.output_pre = distribution.sample
# 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(distribution.log_probs,
name="action_probs")
self.entropy = distribution.entropy
# We keep these tensors the same name, but use new nodes to keep code parallelism with discrete control.
self.total_log_probs = distribution.total_log_probs
def normalize_vector_obs(self, vector_obs):
normalized_state = tf.clip_by_value(
(vector_obs - self.running_mean) /
tf.sqrt(self.running_variance /
(tf.cast(self.normalization_steps, tf.float32) + 1)),
-5,
5,
name="normalized_state",
)
return normalized_state
def normalize_vector_obs(
vector_obs: tf.Tensor,
running_mean: tf.Tensor,
running_variance: tf.Tensor,
normalization_steps: tf.Tensor,
) -> tf.Tensor:
"""
Create a normalized version of an input tensor.
:param vector_obs: Input vector observation tensor.
:param running_mean: Tensorflow tensor representing the current running mean.
:param running_variance: Tensorflow tensor representing the current running variance.
:param normalization_steps: Tensorflow tensor representing the current number of normalization_steps.
:return: A normalized version of vector_obs.
"""
normalized_state = tf.clip_by_value(
(vector_obs - running_mean) /
tf.sqrt(running_variance /
(tf.cast(normalization_steps, tf.float32) + 1)),
-5,
5,
name="normalized_state",
)
return normalized_state
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_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 __init__(
self,
brain,
h_size=128,
lr=1e-4,
n_layers=2,
m_size=128,
normalize=False,
use_recurrent=False,
seed=0,
):
LearningModel.__init__(self, m_size, normalize, use_recurrent, brain, seed)
num_streams = 1
hidden_streams = self.create_observation_streams(num_streams, h_size, n_layers)
hidden = hidden_streams[0]
self.dropout_rate = tf.placeholder(
dtype=tf.float32, shape=[], name="dropout_rate"
)
hidden_reg = tf.layers.dropout(hidden, self.dropout_rate)
if self.use_recurrent:
tf.Variable(
self.m_size, name="memory_size", trainable=False, dtype=tf.int32
)
self.memory_in = tf.placeholder(
shape=[None, self.m_size], dtype=tf.float32, name="recurrent_in"
)
hidden_reg, self.memory_out = self.create_recurrent_encoder(
hidden_reg, self.memory_in, self.sequence_length
)
self.memory_out = tf.identity(self.memory_out, name="recurrent_out")
if brain.vector_action_space_type == "discrete":
policy_branches = []
for size in self.act_size:
policy_branches.append(
tf.layers.dense(
hidden_reg,
size,
activation=None,
use_bias=False,
kernel_initializer=tf.initializers.variance_scaling(0.01),
)
)
self.action_probs = tf.concat(
[tf.nn.softmax(branch) for branch in policy_branches],
axis=1,
name="action_probs",
)
self.action_masks = tf.placeholder(
shape=[None, sum(self.act_size)], dtype=tf.float32, name="action_masks"
)
self.sample_action_float, _, normalized_logits = self.create_discrete_action_masking_layer(
tf.concat(policy_branches, axis=1), self.action_masks, self.act_size
)
tf.identity(normalized_logits, name="action")
self.sample_action = tf.cast(self.sample_action_float, tf.int32)
self.true_action = tf.placeholder(
shape=[None, len(policy_branches)],
dtype=tf.int32,
name="teacher_action",
)
self.action_oh = tf.concat(
[
tf.one_hot(self.true_action[:, i], self.act_size[i])
for i in range(len(self.act_size))
],
axis=1,
)
self.loss = tf.reduce_sum(
-tf.log(self.action_probs + 1e-10) * self.action_oh
)
self.action_percent = tf.reduce_mean(
tf.cast(
tf.equal(
tf.cast(tf.argmax(self.action_probs, axis=1), tf.int32),
self.sample_action,
),
tf.float32,
)
)
else:
self.policy = tf.layers.dense(
hidden_reg,
self.act_size[0],
activation=None,
use_bias=False,
name="pre_action",
kernel_initializer=tf.initializers.variance_scaling(0.01),
)
self.clipped_sample_action = tf.clip_by_value(self.policy, -1, 1)
self.sample_action = tf.identity(self.clipped_sample_action, name="action")
self.true_action = tf.placeholder(
shape=[None, self.act_size[0]], dtype=tf.float32, name="teacher_action"
)
self.clipped_true_action = tf.clip_by_value(self.true_action, -1, 1)
self.loss = tf.reduce_sum(
tf.squared_difference(self.clipped_true_action, self.sample_action)
)
optimizer = tf.train.AdamOptimizer(learning_rate=lr)
self.update = optimizer.minimize(self.loss)
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")