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 make_inputs(self) -> None:
"""
Creates the input layers for the discriminator
"""
self.done_expert_holder = tf.placeholder(shape=[None], dtype=tf.float32)
self.done_policy_holder = tf.placeholder(shape=[None], dtype=tf.float32)
self.done_expert = tf.expand_dims(self.done_expert_holder, -1)
self.done_policy = tf.expand_dims(self.done_policy_holder, -1)
if self.policy.brain.vector_action_space_type == "continuous":
action_length = self.policy.act_size[0]
self.action_in_expert = tf.placeholder(
shape=[None, action_length], dtype=tf.float32
)
self.expert_action = tf.identity(self.action_in_expert)
else:
action_length = len(self.policy.act_size)
self.action_in_expert = tf.placeholder(
shape=[None, action_length], dtype=tf.int32
)
self.expert_action = tf.concat(
[
tf.one_hot(self.action_in_expert[:, i], act_size)
for i, act_size in enumerate(self.policy.act_size)
],
axis=1,
)
encoded_policy_list = []
encoded_expert_list = []
if self.policy.vec_obs_size > 0:
self.obs_in_expert = tf.placeholder(
shape=[None, self.policy.vec_obs_size], dtype=tf.float32
)
if self.policy.normalize:
encoded_expert_list.append(
ModelUtils.normalize_vector_obs(
self.obs_in_expert,
self.policy.running_mean,
self.policy.running_variance,
self.policy.normalization_steps,
)
)
encoded_policy_list.append(self.policy.processed_vector_in)
else:
encoded_expert_list.append(self.obs_in_expert)
encoded_policy_list.append(self.policy.vector_in)
if self.policy.vis_obs_size > 0:
self.expert_visual_in: List[tf.Tensor] = []
visual_policy_encoders = []
visual_expert_encoders = []
for i in range(self.policy.vis_obs_size):
# Create input ops for next (t+1) visual observations.
visual_input = ModelUtils.create_visual_input(
self.policy.brain.camera_resolutions[i],
name="gail_visual_observation_" + str(i),
)
self.expert_visual_in.append(visual_input)
encoded_policy_visual = ModelUtils.create_visual_observation_encoder(
self.policy.visual_in[i],
self.encoding_size,
ModelUtils.swish,
1,
"gail_stream_{}_visual_obs_encoder".format(i),
False,
)
encoded_expert_visual = ModelUtils.create_visual_observation_encoder(
self.expert_visual_in[i],
self.encoding_size,
ModelUtils.swish,
1,
"gail_stream_{}_visual_obs_encoder".format(i),
True,
)
visual_policy_encoders.append(encoded_policy_visual)
visual_expert_encoders.append(encoded_expert_visual)
hidden_policy_visual = tf.concat(visual_policy_encoders, axis=1)
hidden_expert_visual = tf.concat(visual_expert_encoders, axis=1)
encoded_policy_list.append(hidden_policy_visual)
encoded_expert_list.append(hidden_expert_visual)
self.encoded_expert = tf.concat(encoded_expert_list, axis=1)
self.encoded_policy = tf.concat(encoded_policy_list, axis=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 make_inputs(self) -> None:
"""
Creates the input layers for the discriminator
"""
self.done_expert_holder = tf.placeholder(shape=[None], dtype=tf.float32)
self.done_policy_holder = tf.placeholder(shape=[None], dtype=tf.float32)
self.done_expert = tf.expand_dims(self.done_expert_holder, -1)
self.done_policy = tf.expand_dims(self.done_policy_holder, -1)
if self.policy.behavior_spec.is_action_continuous():
action_length = self.policy.act_size[0]
self.action_in_expert = tf.placeholder(
shape=[None, action_length], dtype=tf.float32
)
self.expert_action = tf.identity(self.action_in_expert)
else:
action_length = len(self.policy.act_size)
self.action_in_expert = tf.placeholder(
shape=[None, action_length], dtype=tf.int32
)
self.expert_action = tf.concat(
[
tf.one_hot(self.action_in_expert[:, i], act_size)
for i, act_size in enumerate(self.policy.act_size)
],
axis=1,
)
encoded_policy_list = []
encoded_expert_list = []
(
self.obs_in_expert,
self.expert_visual_in,
) = ModelUtils.create_input_placeholders(
self.policy.behavior_spec.observation_shapes, "gail_"
)
if self.policy.vec_obs_size > 0:
if self.policy.normalize:
encoded_expert_list.append(
ModelUtils.normalize_vector_obs(
self.obs_in_expert,
self.policy.running_mean,
self.policy.running_variance,
self.policy.normalization_steps,
)
)
encoded_policy_list.append(self.policy.processed_vector_in)
else:
encoded_expert_list.append(self.obs_in_expert)
encoded_policy_list.append(self.policy.vector_in)
if self.expert_visual_in:
visual_policy_encoders = []
visual_expert_encoders = []
for i, (vis_in, exp_vis_in) in enumerate(
zip(self.policy.visual_in, self.expert_visual_in)
):
encoded_policy_visual = ModelUtils.create_visual_observation_encoder(
vis_in,
self.encoding_size,
ModelUtils.swish,
1,
f"gail_stream_{i}_visual_obs_encoder",
False,
)
encoded_expert_visual = ModelUtils.create_visual_observation_encoder(
exp_vis_in,
self.encoding_size,
ModelUtils.swish,
1,
f"gail_stream_{i}_visual_obs_encoder",
True,
)
visual_policy_encoders.append(encoded_policy_visual)
visual_expert_encoders.append(encoded_expert_visual)
hidden_policy_visual = tf.concat(visual_policy_encoders, axis=1)
hidden_expert_visual = tf.concat(visual_expert_encoders, axis=1)
encoded_policy_list.append(hidden_policy_visual)
encoded_expert_list.append(hidden_expert_visual)
self.encoded_expert = tf.concat(encoded_expert_list, axis=1)
self.encoded_policy = tf.concat(encoded_policy_list, axis=1)
