def _action_onehot(self, sample: tf.Tensor,
act_size: List[int]) -> tf.Tensor:
action_oh = tf.concat(
[
tf.one_hot(sample[:, i], act_size[i])
for i in range(len(act_size))
],
axis=1,
)
return action_oh
def _create_dc_actor(self, encoded: tf.Tensor) -> None:
"""
Creates Discrete 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.
"""
if self.use_recurrent:
self.prev_action = tf.placeholder(shape=[None,
len(self.act_size)],
dtype=tf.int32,
name="prev_action")
prev_action_oh = tf.concat(
[
tf.one_hot(self.prev_action[:, i], self.act_size[i])
for i in range(len(self.act_size))
],
axis=1,
)
hidden_policy = tf.concat([encoded, prev_action_oh], axis=1)
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(
hidden_policy,
self.memory_in,
self.sequence_length_ph,
name="lstm_policy",
)
self.memory_out = tf.identity(memory_policy_out, "recurrent_out")
else:
hidden_policy = encoded
self.action_masks = tf.placeholder(shape=[None,
sum(self.act_size)],
dtype=tf.float32,
name="action_masks")
with tf.variable_scope("policy"):
distribution = MultiCategoricalDistribution(
hidden_policy, self.act_size, self.action_masks)
# It's important that we are able to feed_dict a value into this tensor to get the
# right one-hot encoding, so we can't do identity on it.
self.output = distribution.sample
self.all_log_probs = tf.identity(distribution.log_probs, name="action")
self.selected_actions = tf.stop_gradient(
distribution.sample_onehot) # In discrete, these are onehot
self.entropy = distribution.entropy
self.total_log_probs = distribution.total_log_probs
def make_inputs(self) -> None:
"""
Creates the input layers for the discriminator
"""
self.done_expert = tf.placeholder(shape=[None, 1], dtype=tf.float32)
self.done_policy = tf.placeholder(shape=[None, 1], dtype=tf.float32)
if self.policy.behavior_spec.action_spec.is_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,
)
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_dc_actor(self, hidden_policy, scope):
"""
Creates Discrete 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.
"""
scope = self.join_scopes(scope, "policy")
# Create inputs outside of the scope
self.action_masks = tf.placeholder(shape=[None,
sum(self.act_size)],
dtype=tf.float32,
name="action_masks")
if self.use_recurrent:
self.prev_action = tf.placeholder(shape=[None,
len(self.act_size)],
dtype=tf.int32,
name="prev_action")
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:
prev_action_oh = tf.concat(
[
tf.one_hot(self.prev_action[:, i], self.act_size[i])
for i in range(len(self.act_size))
],
axis=1,
)
hidden_policy = tf.concat([hidden_policy, prev_action_oh], axis=1)
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):
policy_branches = []
for size in self.act_size:
policy_branches.append(
tf.layers.dense(
hidden_policy,
size,
activation=None,
use_bias=False,
kernel_initializer=tf.initializers.variance_scaling(
0.01),
))
all_logits = tf.concat(policy_branches,
axis=1,
name="action_probs")
output, normalized_probs, normalized_logprobs = self.create_discrete_action_masking_layer(
all_logits, self.action_masks, self.act_size)
self.action_probs = normalized_probs
# Really, this is entropy, but it has an analogous purpose to the log probs in the
# continuous case.
self.all_log_probs = self.action_probs * normalized_logprobs
self.output = output
# Create action input (discrete)
self.action_holder = tf.placeholder(
shape=[None, len(policy_branches)],
dtype=tf.int32,
name="action_holder")
self.output_oh = tf.concat(
[
tf.one_hot(self.action_holder[:, i], self.act_size[i])
for i in range(len(self.act_size))
],
axis=1,
)
# For Curiosity and GAIL to retrieve selected actions. We don't
# need the mask at this point because it's already stored in the buffer.
self.selected_actions = tf.stop_gradient(self.output_oh)
self.external_action_in = tf.concat(
[
tf.one_hot(self.action_holder[:, i], self.act_size[i])
for i in range(len(self.act_size))
],
axis=1,
)
# This is total entropy over all branches
self.entropy = -1 * tf.reduce_sum(self.all_log_probs, axis=1)
# Extract the normalized logprobs for Barracuda
self.normalized_logprobs = tf.identity(normalized_logprobs,
name="action")
# We kept the LSTMs at a different scope than the rest, so add them if they exist.
self.policy_vars = self.get_vars(scope)
if self.use_recurrent:
self.policy_vars += self.get_vars("lstm")
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)
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_dc_actor_critic(self, h_size: int, num_layers: int,
vis_encode_type: EncoderType) -> None:
"""
Creates Discrete 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(
1, h_size, num_layers, vis_encode_type)
hidden = hidden_streams[0]
if self.use_recurrent:
self.prev_action = tf.placeholder(shape=[None,
len(self.act_size)],
dtype=tf.int32,
name="prev_action")
prev_action_oh = tf.concat(
[
tf.one_hot(self.prev_action[:, i], self.act_size[i])
for i in range(len(self.act_size))
],
axis=1,
)
hidden = tf.concat([hidden, prev_action_oh], axis=1)
self.memory_in = tf.placeholder(shape=[None, self.m_size],
dtype=tf.float32,
name="recurrent_in")
hidden, memory_out = self.create_recurrent_encoder(
hidden, self.memory_in, self.sequence_length)
self.memory_out = tf.identity(memory_out, name="recurrent_out")
policy_branches = []
for size in self.act_size:
policy_branches.append(
tf.layers.dense(
hidden,
size,
activation=None,
use_bias=False,
kernel_initializer=LearningModel.scaled_init(0.01),
))
self.all_log_probs = tf.concat(policy_branches,
axis=1,
name="action_probs")
self.action_masks = tf.placeholder(shape=[None,
sum(self.act_size)],
dtype=tf.float32,
name="action_masks")
output, _, normalized_logits = self.create_discrete_action_masking_layer(
self.all_log_probs, self.action_masks, self.act_size)
self.output = tf.identity(output)
self.normalized_logits = tf.identity(normalized_logits, name="action")
self.create_value_heads(self.stream_names, hidden)
self.action_holder = tf.placeholder(shape=[None,
len(policy_branches)],
dtype=tf.int32,
name="action_holder")
self.action_oh = tf.concat(
[
tf.one_hot(self.action_holder[:, i], self.act_size[i])
for i in range(len(self.act_size))
],
axis=1,
)
self.selected_actions = tf.stop_gradient(self.action_oh)
self.all_old_log_probs = tf.placeholder(
shape=[None, sum(self.act_size)],
dtype=tf.float32,
name="old_probabilities")
_, _, old_normalized_logits = self.create_discrete_action_masking_layer(
self.all_old_log_probs, self.action_masks, self.act_size)
action_idx = [0] + list(np.cumsum(self.act_size))
self.entropy = tf.reduce_sum(
(tf.stack(
[
tf.nn.softmax_cross_entropy_with_logits_v2(
labels=tf.nn.softmax(
self.all_log_probs[:,
action_idx[i]:action_idx[i +
1]]),
logits=self.all_log_probs[:,
action_idx[i]:action_idx[i +
1]],
) for i in range(len(self.act_size))
],
axis=1,
)),
axis=1,
)
self.log_probs = tf.reduce_sum(
(tf.stack(
[
-tf.nn.softmax_cross_entropy_with_logits_v2(
labels=self.action_oh[:,
action_idx[i]:action_idx[i + 1]],
logits=normalized_logits[:,
action_idx[i]:action_idx[i +
1]],
) for i in range(len(self.act_size))
],
axis=1,
)),
axis=1,
keepdims=True,
)
self.old_log_probs = tf.reduce_sum(
(tf.stack(
[
-tf.nn.softmax_cross_entropy_with_logits_v2(
labels=self.action_oh[:,
action_idx[i]:action_idx[i + 1]],
logits=old_normalized_logits[:, action_idx[i]:
action_idx[i + 1]],
) for i in range(len(self.act_size))
],
axis=1,
)),
axis=1,
keepdims=True,
)
def _create_dc_actor(self, encoded: tf.Tensor) -> None:
"""
Creates Discrete 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.
"""
if self.use_recurrent:
self.prev_action = tf.placeholder(shape=[None,
len(self.act_size)],
dtype=tf.int32,
name="prev_action")
prev_action_oh = tf.concat(
[
tf.one_hot(self.prev_action[:, i], self.act_size[i])
for i in range(len(self.act_size))
],
axis=1,
)
hidden_policy = tf.concat([encoded, prev_action_oh], axis=1)
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(
hidden_policy,
self.memory_in,
self.sequence_length_ph,
name="lstm_policy",
)
self.memory_out = tf.identity(memory_policy_out, "recurrent_out")
else:
hidden_policy = encoded
policy_branches = []
with tf.variable_scope("policy"):
for size in self.act_size:
policy_branches.append(
tf.layers.dense(
hidden_policy,
size,
activation=None,
use_bias=False,
kernel_initializer=ModelUtils.scaled_init(0.01),
))
raw_log_probs = tf.concat(policy_branches, axis=1, name="action_probs")
self.action_masks = tf.placeholder(shape=[None,
sum(self.act_size)],
dtype=tf.float32,
name="action_masks")
output, self.action_probs, normalized_logits = ModelUtils.create_discrete_action_masking_layer(
raw_log_probs, self.action_masks, self.act_size)
self.output = tf.identity(output)
self.all_log_probs = tf.identity(normalized_logits, name="action")
self.action_holder = tf.placeholder(shape=[None,
len(policy_branches)],
dtype=tf.int32,
name="action_holder")
self.action_oh = tf.concat(
[
tf.one_hot(self.action_holder[:, i], self.act_size[i])
for i in range(len(self.act_size))
],
axis=1,
)
self.selected_actions = tf.stop_gradient(self.action_oh)
action_idx = [0] + list(np.cumsum(self.act_size))
self.entropy = tf.reduce_sum(
(tf.stack(
[
tf.nn.softmax_cross_entropy_with_logits_v2(
labels=tf.nn.softmax(
self.all_log_probs[:,
action_idx[i]:action_idx[i +
1]]),
logits=self.all_log_probs[:,
action_idx[i]:action_idx[i +
1]],
) for i in range(len(self.act_size))
],
axis=1,
)),
axis=1,
)
self.log_probs = tf.reduce_sum(
(tf.stack(
[
-tf.nn.softmax_cross_entropy_with_logits_v2(
labels=self.action_oh[:,
action_idx[i]:action_idx[i + 1]],
logits=normalized_logits[:,
action_idx[i]:action_idx[i +
1]],
) for i in range(len(self.act_size))
],
axis=1,
)),
axis=1,
keepdims=True,
)
