def __init__(
self, m_size, normalize, use_recurrent, brain, seed, stream_names=None
):
tf.set_random_seed(seed)
self.brain = brain
self.vector_in = None
self.global_step, self.increment_step, self.steps_to_increment = (
self.create_global_steps()
)
self.visual_in = []
self.batch_size = tf.placeholder(shape=None, dtype=tf.int32, name="batch_size")
self.sequence_length = tf.placeholder(
shape=None, dtype=tf.int32, name="sequence_length"
)
self.mask_input = tf.placeholder(shape=[None], dtype=tf.float32, name="masks")
self.mask = tf.cast(self.mask_input, tf.int32)
self.stream_names = stream_names or []
self.use_recurrent = use_recurrent
if self.use_recurrent:
self.m_size = m_size
else:
self.m_size = 0
self.normalize = normalize
self.act_size = brain.vector_action_space_size
self.vec_obs_size = brain.vector_observation_space_size
self.vis_obs_size = brain.number_visual_observations
tf.Variable(
int(brain.vector_action_space_type == "continuous"),
name="is_continuous_control",
trainable=False,
dtype=tf.int32,
)
tf.Variable(
self._version_number_,
name="version_number",
trainable=False,
dtype=tf.int32,
)
tf.Variable(self.m_size, name="memory_size", trainable=False, dtype=tf.int32)
if brain.vector_action_space_type == "continuous":
tf.Variable(
self.act_size[0],
name="action_output_shape",
trainable=False,
dtype=tf.int32,
)
else:
tf.Variable(
sum(self.act_size),
name="action_output_shape",
trainable=False,
dtype=tf.int32,
)
self.value_heads: Dict[str, tf.Tensor] = {}
self.normalization_steps: Optional[tf.Variable] = None
self.running_mean: Optional[tf.Variable] = None
self.running_variance: Optional[tf.Variable] = None
self.update_normalization: Optional[tf.Operation] = None
self.value: Optional[tf.Tensor] = None
def test_average_gradients(mock_get_devices, dummy_config):
tf.reset_default_graph()
mock_get_devices.return_value = [
"/device:GPU:0",
"/device:GPU:1",
"/device:GPU:2",
"/device:GPU:3",
]
trainer_parameters = dummy_config
trainer_parameters["model_path"] = ""
trainer_parameters["keep_checkpoints"] = 3
brain = create_mock_brainparams()
with tf.Session() as sess:
policy = MultiGpuPPOPolicy(0, brain, trainer_parameters, False, False)
var = tf.Variable(0)
tower_grads = [
[(tf.constant(0.1), var)],
[(tf.constant(0.2), var)],
[(tf.constant(0.3), var)],
[(tf.constant(0.4), var)],
]
avg_grads = policy.average_gradients(tower_grads)
init = tf.global_variables_initializer()
sess.run(init)
run_out = sess.run(avg_grads)
assert run_out == [(0.25, 0)]
def test_tanh_distribution():
with tf.Graph().as_default():
logits = tf.Variable(initial_value=[[0, 0]],
trainable=True,
dtype=tf.float32)
distribution = GaussianDistribution(logits,
act_size=VECTOR_ACTION_SPACE,
reparameterize=False,
tanh_squash=True)
sess = tf.Session()
with tf.Session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
output = sess.run(distribution.sample)
for _ in range(10):
output = sess.run(
[distribution.sample, distribution.log_probs])
for out in output:
assert out.shape[1] == VECTOR_ACTION_SPACE[0]
# Assert action never exceeds [-1,1]
action = output[0][0]
for act in action:
assert act >= -1 and act <= 1
output = sess.run([distribution.total_log_probs])
assert output[0].shape[0] == 1
def create_schedule(
schedule: ScheduleType,
parameter: float,
global_step: tf.Tensor,
max_step: int,
min_value: float,
) -> tf.Tensor:
"""
Create a learning rate tensor.
:param lr_schedule: Type of learning rate schedule.
:param lr: Base learning rate.
:param global_step: A TF Tensor representing the total global step.
:param max_step: The maximum number of steps in the training run.
:return: A Tensor containing the learning rate.
"""
if schedule == ScheduleType.CONSTANT:
parameter_rate = tf.Variable(parameter, trainable=False)
elif schedule == ScheduleType.LINEAR:
parameter_rate = tf.train.polynomial_decay(parameter,
global_step,
max_step,
min_value,
power=1.0)
else:
raise UnityTrainerException(f"The schedule {schedule} is invalid.")
return parameter_rate
def test_gaussian_distribution():
with tf.Graph().as_default():
logits = tf.Variable(initial_value=[[1, 1]],
trainable=True,
dtype=tf.float32)
distribution = GaussianDistribution(
logits,
act_size=VECTOR_ACTION_SPACE,
reparameterize=False,
tanh_squash=False,
)
sess = tf.Session()
with tf.Session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
output = sess.run(distribution.sample)
for _ in range(10):
output = sess.run(
[distribution.sample, distribution.log_probs])
for out in output:
assert out.shape[1] == VECTOR_ACTION_SPACE[0]
output = sess.run([distribution.total_log_probs])
assert output[0].shape[0] == 1
# Test entropy is correct
log_std_tensor = tf.get_default_graph().get_tensor_by_name(
"log_std/BiasAdd:0")
feed_dict = {log_std_tensor: [[1.0, 1.0]]}
entropy = sess.run([distribution.entropy], feed_dict=feed_dict)
# Entropy with log_std of 1.0 should be 2.42
assert pytest.approx(entropy[0], 0.01) == 2.42
def create_loss(self, learning_rate: float, anneal_steps: int) -> None:
"""
Creates the loss and update nodes for the BC module
:param learning_rate: The learning rate for the optimizer
:param anneal_steps: Number of steps over which to anneal the learning_rate
"""
selected_action = self.policy.output
if self.policy.use_continuous_act:
self.loss = tf.reduce_mean(
tf.squared_difference(selected_action, self.expert_action))
else:
log_probs = self.policy.all_log_probs
self.loss = tf.reduce_mean(
-tf.log(tf.nn.softmax(log_probs) + 1e-7) * self.expert_action)
if anneal_steps > 0:
self.annealed_learning_rate = tf.train.polynomial_decay(
learning_rate,
self.policy.global_step,
anneal_steps,
0.0,
power=1.0)
else:
self.annealed_learning_rate = tf.Variable(learning_rate)
optimizer = tf.train.AdamOptimizer(
learning_rate=self.annealed_learning_rate, name="bc_adam")
self.update_batch = optimizer.minimize(self.loss)
def create_learning_rate(
lr_schedule: LearningRateSchedule,
lr: float,
global_step: tf.Tensor,
max_step: int,
) -> tf.Tensor:
"""
Create a learning rate tensor.
:param lr_schedule: Type of learning rate schedule.
:param lr: Base learning rate.
:param global_step: A TF Tensor representing the total global step.
:param max_step: The maximum number of steps in the training run.
:return: A Tensor containing the learning rate.
"""
if lr_schedule == LearningRateSchedule.CONSTANT:
learning_rate = tf.Variable(lr)
elif lr_schedule == LearningRateSchedule.LINEAR:
learning_rate = tf.train.polynomial_decay(lr,
global_step,
max_step,
1e-10,
power=1.0)
else:
raise UnityTrainerException(
"The learning rate schedule {} is invalid.".format(
lr_schedule))
return learning_rate
def test_multicategorical_distribution():
with tf.Graph().as_default():
logits = tf.Variable(initial_value=[[0, 0]],
trainable=True,
dtype=tf.float32)
action_masks = tf.Variable(
initial_value=[[1 for _ in range(sum(DISCRETE_ACTION_SPACE))]],
trainable=True,
dtype=tf.float32,
)
distribution = MultiCategoricalDistribution(
logits, act_size=DISCRETE_ACTION_SPACE, action_masks=action_masks)
sess = tf.Session()
with tf.Session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
output = sess.run(distribution.sample)
for _ in range(10):
sample, log_probs, entropy = sess.run([
distribution.sample, distribution.log_probs,
distribution.entropy
])
assert len(log_probs[0]) == sum(DISCRETE_ACTION_SPACE)
# Assert action never exceeds [-1,1]
assert len(sample[0]) == len(DISCRETE_ACTION_SPACE)
for i, act in enumerate(sample[0]):
assert act >= 0 and act <= DISCRETE_ACTION_SPACE[i]
output = sess.run([distribution.total_log_probs])
assert output[0].shape[0] == 1
# Make sure entropy is correct
assert entropy[0] > 3.8
# Test masks
mask = []
for space in DISCRETE_ACTION_SPACE:
mask.append(1)
for _action_space in range(1, space):
mask.append(0)
for _ in range(10):
sample, log_probs = sess.run(
[distribution.sample, distribution.log_probs],
feed_dict={action_masks: [mask]},
)
for act in sample[0]:
assert act >= 0 and act <= 1
output = sess.run([distribution.total_log_probs])
def create_global_steps():
"""Creates TF ops to track and increment global training step."""
global_step = tf.Variable(
0, name="global_step", trainable=False, dtype=tf.int32
)
steps_to_increment = tf.placeholder(
shape=[], dtype=tf.int32, name="steps_to_increment"
)
increment_step = tf.assign(global_step, tf.add(global_step, steps_to_increment))
return global_step, increment_step, steps_to_increment
def create_learning_rate(
lr_schedule: LearningRateSchedule,
lr: float,
global_step: tf.Tensor,
max_step: int,
) -> tf.Tensor:
if lr_schedule == LearningRateSchedule.CONSTANT:
learning_rate = tf.Variable(lr)
elif lr_schedule == LearningRateSchedule.LINEAR:
learning_rate = tf.train.polynomial_decay(
lr, global_step, max_step, 1e-10, power=1.0
)
else:
raise UnityTrainerException(
"The learning rate schedule {} is invalid.".format(lr_schedule)
)
return learning_rate
def create_input_placeholders(self):
with self.graph.as_default():
(
self.global_step,
self.increment_step_op,
self.steps_to_increment,
) = ModelUtils.create_global_steps()
self.vector_in, self.visual_in = ModelUtils.create_input_placeholders(
self.behavior_spec.observation_shapes)
if self.normalize:
self.first_normalization_update = True
normalization_tensors = ModelUtils.create_normalizer(
self.vector_in)
self.update_normalization_op = normalization_tensors.update_op
self.init_normalization_op = normalization_tensors.init_op
self.normalization_steps = normalization_tensors.steps
self.running_mean = normalization_tensors.running_mean
self.running_variance = normalization_tensors.running_variance
self.processed_vector_in = ModelUtils.normalize_vector_obs(
self.vector_in,
self.running_mean,
self.running_variance,
self.normalization_steps,
)
else:
self.processed_vector_in = self.vector_in
self.update_normalization_op = None
self.batch_size_ph = tf.placeholder(shape=None,
dtype=tf.int32,
name="batch_size")
self.sequence_length_ph = tf.placeholder(shape=None,
dtype=tf.int32,
name="sequence_length")
self.mask_input = tf.placeholder(shape=[None],
dtype=tf.float32,
name="masks")
# Only needed for PPO, but needed for BC module
self.epsilon = tf.placeholder(shape=[None, self.act_size[0]],
dtype=tf.float32,
name="epsilon")
self.mask = tf.cast(self.mask_input, tf.int32)
tf.Variable(
int(self.behavior_spec.is_action_continuous()),
name="is_continuous_control",
trainable=False,
dtype=tf.int32,
)
int_version = TFPolicy._convert_version_string(__version__)
major_ver_t = tf.Variable(
int_version[0],
name="trainer_major_version",
trainable=False,
dtype=tf.int32,
)
minor_ver_t = tf.Variable(
int_version[1],
name="trainer_minor_version",
trainable=False,
dtype=tf.int32,
)
patch_ver_t = tf.Variable(
int_version[2],
name="trainer_patch_version",
trainable=False,
dtype=tf.int32,
)
self.version_tensors = (major_ver_t, minor_ver_t, patch_ver_t)
tf.Variable(
MODEL_FORMAT_VERSION,
name="version_number",
trainable=False,
dtype=tf.int32,
)
tf.Variable(self.m_size,
name="memory_size",
trainable=False,
dtype=tf.int32)
if self.behavior_spec.is_action_continuous():
tf.Variable(
self.act_size[0],
name="action_output_shape",
trainable=False,
dtype=tf.int32,
)
else:
tf.Variable(
sum(self.act_size),
name="action_output_shape",
trainable=False,
dtype=tf.int32,
)
def __init__(self, policy: TFPolicy, trainer_params: TrainerSettings):
"""
Takes a Unity environment and model-specific hyper-parameters and returns the
appropriate PPO agent model for the environment.
:param brain: Brain parameters used to generate specific network graph.
:param lr: Learning rate.
:param lr_schedule: Learning rate decay schedule.
:param h_size: Size of hidden layers
:param init_entcoef: Initial value for entropy coefficient. Set lower to learn faster,
set higher to explore more.
:return: a sub-class of PPOAgent tailored to the environment.
:param max_step: Total number of training steps.
:param normalize: Whether to normalize vector observation input.
:param use_recurrent: Whether to use an LSTM layer in the network.
:param num_layers: Number of hidden layers between encoded input and policy & value layers
:param tau: Strength of soft-Q update.
:param m_size: Size of brain memory.
"""
# Create the graph here to give more granular control of the TF graph to the Optimizer.
policy.create_tf_graph()
with policy.graph.as_default():
with tf.variable_scope(""):
super().__init__(policy, trainer_params)
hyperparameters: SACSettings = cast(
SACSettings, trainer_params.hyperparameters)
lr = hyperparameters.learning_rate
lr_schedule = hyperparameters.learning_rate_schedule
max_step = trainer_params.max_steps
self.tau = hyperparameters.tau
self.init_entcoef = hyperparameters.init_entcoef
self.policy = policy
self.act_size = policy.act_size
policy_network_settings = policy.network_settings
h_size = policy_network_settings.hidden_units
num_layers = policy_network_settings.num_layers
vis_encode_type = policy_network_settings.vis_encode_type
self.tau = hyperparameters.tau
self.burn_in_ratio = 0.0
# Non-exposed SAC parameters
self.discrete_target_entropy_scale = (
0.2 # Roughly equal to e-greedy 0.05
)
self.continuous_target_entropy_scale = 1.0
stream_names = list(self.reward_signals.keys())
# Use to reduce "survivor bonus" when using Curiosity or GAIL.
self.gammas = [
_val.gamma
for _val in trainer_params.reward_signals.values()
]
self.use_dones_in_backup = {
name: tf.Variable(1.0)
for name in stream_names
}
self.disable_use_dones = {
name: self.use_dones_in_backup[name].assign(0.0)
for name in stream_names
}
if num_layers < 1:
num_layers = 1
self.target_init_op: List[tf.Tensor] = []
self.target_update_op: List[tf.Tensor] = []
self.update_batch_policy: Optional[tf.Operation] = None
self.update_batch_value: Optional[tf.Operation] = None
self.update_batch_entropy: Optional[tf.Operation] = None
self.policy_network = SACPolicyNetwork(
policy=self.policy,
m_size=self.policy.m_size, # 3x policy.m_size
h_size=h_size,
normalize=self.policy.normalize,
use_recurrent=self.policy.use_recurrent,
num_layers=num_layers,
stream_names=stream_names,
vis_encode_type=vis_encode_type,
)
self.target_network = SACTargetNetwork(
policy=self.policy,
m_size=self.policy.m_size, # 1x policy.m_size
h_size=h_size,
normalize=self.policy.normalize,
use_recurrent=self.policy.use_recurrent,
num_layers=num_layers,
stream_names=stream_names,
vis_encode_type=vis_encode_type,
)
# The optimizer's m_size is 3 times the policy (Q1, Q2, and Value)
self.m_size = 3 * self.policy.m_size
self._create_inputs_and_outputs()
self.learning_rate = ModelUtils.create_schedule(
lr_schedule,
lr,
self.policy.global_step,
int(max_step),
min_value=1e-10,
)
self._create_losses(
self.policy_network.q1_heads,
self.policy_network.q2_heads,
lr,
int(max_step),
stream_names,
discrete=not self.policy.use_continuous_act,
)
self._create_sac_optimizer_ops()
self.selected_actions = (self.policy.selected_actions
) # For GAIL and other reward signals
if self.policy.normalize:
target_update_norm = self.target_network.copy_normalization(
self.policy.running_mean,
self.policy.running_variance,
self.policy.normalization_steps,
)
# Update the normalization of the optimizer when the policy does.
self.policy.update_normalization_op = tf.group([
self.policy.update_normalization_op, target_update_norm
])
self.stats_name_to_update_name = {
"Losses/Value Loss": "value_loss",
"Losses/Policy Loss": "policy_loss",
"Losses/Q1 Loss": "q1_loss",
"Losses/Q2 Loss": "q2_loss",
"Policy/Entropy Coeff": "entropy_coef",
"Policy/Learning Rate": "learning_rate",
}
self.update_dict = {
"value_loss": self.total_value_loss,
"policy_loss": self.policy_loss,
"q1_loss": self.q1_loss,
"q2_loss": self.q2_loss,
"entropy_coef": self.ent_coef,
"update_batch": self.update_batch_policy,
"update_value": self.update_batch_value,
"update_entropy": self.update_batch_entropy,
"learning_rate": self.learning_rate,
}
def create_input_placeholders(self):
with self.graph.as_default():
(
self.global_step,
self.increment_step_op,
self.steps_to_increment,
) = ModelUtils.create_global_steps()
self.visual_in = ModelUtils.create_visual_input_placeholders(
self.brain.camera_resolutions
)
self.vector_in = ModelUtils.create_vector_input(self.vec_obs_size)
if self.normalize:
normalization_tensors = ModelUtils.create_normalizer(self.vector_in)
self.update_normalization_op = normalization_tensors.update_op
self.normalization_steps = normalization_tensors.steps
self.running_mean = normalization_tensors.running_mean
self.running_variance = normalization_tensors.running_variance
self.processed_vector_in = ModelUtils.normalize_vector_obs(
self.vector_in,
self.running_mean,
self.running_variance,
self.normalization_steps,
)
else:
self.processed_vector_in = self.vector_in
self.update_normalization_op = None
self.batch_size_ph = tf.placeholder(
shape=None, dtype=tf.int32, name="batch_size"
)
self.sequence_length_ph = tf.placeholder(
shape=None, dtype=tf.int32, name="sequence_length"
)
self.mask_input = tf.placeholder(
shape=[None], dtype=tf.float32, name="masks"
)
# Only needed for PPO, but needed for BC module
self.epsilon = tf.placeholder(
shape=[None, self.act_size[0]], dtype=tf.float32, name="epsilon"
)
self.mask = tf.cast(self.mask_input, tf.int32)
tf.Variable(
int(self.brain.vector_action_space_type == "continuous"),
name="is_continuous_control",
trainable=False,
dtype=tf.int32,
)
tf.Variable(
self._version_number_,
name="version_number",
trainable=False,
dtype=tf.int32,
)
tf.Variable(
self.m_size, name="memory_size", trainable=False, dtype=tf.int32
)
if self.brain.vector_action_space_type == "continuous":
tf.Variable(
self.act_size[0],
name="action_output_shape",
trainable=False,
dtype=tf.int32,
)
else:
tf.Variable(
sum(self.act_size),
name="action_output_shape",
trainable=False,
dtype=tf.int32,
)
def __init__(
self,
brain,
lr=1e-4,
lr_schedule=LearningRateSchedule.CONSTANT,
h_size=128,
init_entcoef=0.1,
max_step=5e6,
normalize=False,
use_recurrent=False,
num_layers=2,
m_size=None,
seed=0,
stream_names=None,
tau=0.005,
gammas=None,
vis_encode_type=EncoderType.SIMPLE,
):
"""
Takes a Unity environment and model-specific hyper-parameters and returns the
appropriate PPO agent model for the environment.
:param brain: BrainInfo used to generate specific network graph.
:param lr: Learning rate.
:param lr_schedule: Learning rate decay schedule.
:param h_size: Size of hidden layers
:param init_entcoef: Initial value for entropy coefficient. Set lower to learn faster,
set higher to explore more.
:return: a sub-class of PPOAgent tailored to the environment.
:param max_step: Total number of training steps.
:param normalize: Whether to normalize vector observation input.
:param use_recurrent: Whether to use an LSTM layer in the network.
:param num_layers: Number of hidden layers between encoded input and policy & value layers
:param tau: Strength of soft-Q update.
:param m_size: Size of brain memory.
"""
self.tau = tau
self.gammas = gammas
self.brain = brain
self.init_entcoef = init_entcoef
if stream_names is None:
stream_names = []
# Use to reduce "survivor bonus" when using Curiosity or GAIL.
self.use_dones_in_backup = {
name: tf.Variable(1.0)
for name in stream_names
}
self.disable_use_dones = {
name: self.use_dones_in_backup[name].assign(0.0)
for name in stream_names
}
LearningModel.__init__(self, m_size, normalize, use_recurrent, brain,
seed, stream_names)
if num_layers < 1:
num_layers = 1
self.target_init_op: List[tf.Tensor] = []
self.target_update_op: List[tf.Tensor] = []
self.update_batch_policy: Optional[tf.Operation] = None
self.update_batch_value: Optional[tf.Operation] = None
self.update_batch_entropy: Optional[tf.Operation] = None
self.policy_network = SACPolicyNetwork(
brain=brain,
m_size=m_size,
h_size=h_size,
normalize=normalize,
use_recurrent=use_recurrent,
num_layers=num_layers,
seed=seed,
stream_names=stream_names,
vis_encode_type=vis_encode_type,
)
self.target_network = SACTargetNetwork(
brain=brain,
m_size=m_size // 4 if m_size else None,
h_size=h_size,
normalize=normalize,
use_recurrent=use_recurrent,
num_layers=num_layers,
seed=seed,
stream_names=stream_names,
vis_encode_type=vis_encode_type,
)
self.create_inputs_and_outputs()
self.learning_rate = self.create_learning_rate(lr_schedule, lr,
self.global_step,
max_step)
self.create_losses(
self.policy_network.q1_heads,
self.policy_network.q2_heads,
lr,
max_step,
stream_names,
discrete=self.brain.vector_action_space_type == "discrete",
)
self.selected_actions = (self.policy_network.selected_actions
) # For GAIL and other reward signals
if normalize:
target_update_norm = self.target_network.copy_normalization(
self.policy_network.running_mean,
self.policy_network.running_variance,
self.policy_network.normalization_steps,
)
self.update_normalization = tf.group(
[self.policy_network.update_normalization, target_update_norm])
self.running_mean = self.policy_network.running_mean
self.running_variance = self.policy_network.running_variance
self.normalization_steps = self.policy_network.normalization_steps
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)