def dense_bitwise_categorical_fun(action_space, config, observations):
"""Dense network with bitwise input and categorical output."""
del config
obs_shape = common_layers.shape_list(observations)
x = tf.reshape(observations, [-1] + obs_shape[2:])
with tf.variable_scope("network_parameters"):
with tf.variable_scope("dense_bitwise"):
x = discretization.int_to_bit_embed(x, 8, 32)
flat_x = tf.reshape(x, [
obs_shape[0], obs_shape[1],
functools.reduce(operator.mul,
x.shape.as_list()[1:], 1)
])
x = tf.contrib.layers.fully_connected(flat_x, 256, tf.nn.relu)
x = tf.contrib.layers.fully_connected(flat_x, 128, tf.nn.relu)
logits = tf.contrib.layers.fully_connected(x,
action_space.n,
activation_fn=None)
value = tf.contrib.layers.fully_connected(x, 1,
activation_fn=None)[...,
0]
policy = tf.contrib.distributions.Categorical(logits=logits)
return NetworkOutput(policy, value, lambda a: a)
def bottom(self, x):
inputs = x
with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
common_layers.summarize_video(inputs, "bottom")
# Embed bitwise.
assert self.top_dimensionality == 256
embedded = discretization.int_to_bit_embed(
inputs, 8, self.PIXEL_EMBEDDING_SIZE)
# Project.
return tf.layers.dense(embedded,
self._body_input_depth,
name="merge_pixel_embedded_frames")
def bottom(self, inputs):
with tf.variable_scope(self.name):
common_layers.summarize_video(inputs, "targets_bottom")
# Embed bitwise.
assert self.top_dimensionality == 256
embedded = discretization.int_to_bit_embed(
inputs, 8, self.PIXEL_EMBEDDING_SIZE)
# Transpose and project.
transposed = common_layers.time_to_channels(embedded)
return tf.layers.dense(transposed,
self._body_input_depth,
name="merge_pixel_embedded_frames")
def targets_bottom(self, x): # pylint: disable=arguments-differ
inputs = x
with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
common_layers.summarize_video(inputs, "targets_bottom")
# Embed bitwise.
assert self.top_dimensionality == 256
embedded = discretization.int_to_bit_embed(
inputs, 8, self.PIXEL_EMBEDDING_SIZE)
# Transpose and project.
transposed = common_layers.time_to_channels(embedded)
return tf.layers.dense(transposed,
self._model_hparams.hidden_size,
name="merge_pixel_embedded_frames")
def bottom(self, x):
inputs = x
with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
common_layers.summarize_video(inputs, "bottom")
# Embed bitwise.
assert self._vocab_size == 256
embedded = discretization.int_to_bit_embed(inputs, 8,
self.PIXEL_EMBEDDING_SIZE)
# Project.
return tf.layers.dense(
embedded,
self._model_hparams.hidden_size,
name="merge_pixel_embedded_frames")
def body(self, features):
observations = features["inputs"]
flat_x = tf.layers.flatten(observations)
with tf.variable_scope("dense_bitwise"):
flat_x = discretization.int_to_bit_embed(flat_x, 8, 32)
x = tf.layers.dense(flat_x, 256, activation=tf.nn.relu)
x = tf.layers.dense(flat_x, 128, activation=tf.nn.relu)
logits = tf.layers.dense(x, self.hparams.problem.num_actions)
value = tf.layers.dense(x, 1)[..., 0]
return {"target_policy": logits, "target_value": value}
def targets_bottom(self, x): # pylint: disable=arguments-differ
inputs = x
with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
common_layers.summarize_video(inputs, "targets_bottom")
# Embed bitwise.
assert self.top_dimensionality == 256
embedded = discretization.int_to_bit_embed(inputs, 8,
self.PIXEL_EMBEDDING_SIZE)
# Transpose and project.
transposed = common_layers.time_to_channels(embedded)
return tf.layers.dense(
transposed,
self._body_input_depth,
name="merge_pixel_embedded_frames")
def body(self, features):
observations = features["inputs"]
obs_shape = common_layers.shape_list(observations)
x = tf.reshape(observations, [-1] + obs_shape[2:])
with tf.variable_scope("dense_bitwise"):
x = discretization.int_to_bit_embed(x, 8, 32)
flat_x = tf.reshape(
x, [obs_shape[0], obs_shape[1],
functools.reduce(operator.mul, x.shape.as_list()[1:], 1)])
x = tf.contrib.layers.fully_connected(flat_x, 256, tf.nn.relu)
x = tf.contrib.layers.fully_connected(flat_x, 128, tf.nn.relu)
logits = tf.contrib.layers.fully_connected(
x, self._get_num_actions(features), activation_fn=None
)
value = tf.contrib.layers.fully_connected(
x, 1, activation_fn=None)[..., 0]
return {"target_action": logits, "target_value": value}
def dense_bitwise_categorical_fun(action_space, config, observations):
"""Dense network with bitwise input and categorical output."""
del config
obs_shape = common_layers.shape_list(observations)
x = tf.reshape(observations, [-1] + obs_shape[2:])
with tf.variable_scope("network_parameters"):
with tf.variable_scope("dense_bitwise"):
x = discretization.int_to_bit_embed(x, 8, 32)
flat_x = tf.reshape(
x, [obs_shape[0], obs_shape[1],
functools.reduce(operator.mul, x.shape.as_list()[1:], 1)])
x = tf.contrib.layers.fully_connected(flat_x, 256, tf.nn.relu)
x = tf.contrib.layers.fully_connected(flat_x, 128, tf.nn.relu)
logits = tf.contrib.layers.fully_connected(x, action_space.n,
activation_fn=None)
value = tf.contrib.layers.fully_connected(
x, 1, activation_fn=None)[..., 0]
policy = tfp.distributions.Categorical(logits=logits)
return NetworkOutput(policy, value, lambda a: a)
