def create_agent(sess, environment, summary_writer=None):
"""Creates a DQN agent.
Args:
sess: A `tf.Session` object for running associated ops.
environment: An Atari 2600 Gym environment.
summary_writer: A Tensorflow summary writer to pass to the agent
for in-agent training statistics in Tensorboard.
Returns:
agent: An RL agent.
Raises:
ValueError: If `agent_name` is not in supported list.
"""
if not FLAGS.debug_mode:
summary_writer = None
if FLAGS.agent_name == 'dqn':
return dqn_agent.DQNAgent(sess,
num_actions=5,
summary_writer=summary_writer)
elif FLAGS.agent_name == 'rainbow':
return rainbow_agent.RainbowAgent(sess,
num_actions=5,
summary_writer=summary_writer)
elif FLAGS.agent_name == 'implicit_quantile':
return implicit_quantile_agent.ImplicitQuantileAgent(
sess, num_actions=5, summary_writer=summary_writer)
else:
raise ValueError('Unknown agent: {}'.format(FLAGS.agent_name))
def create_agent(sess,
environment,
agent_name=None,
summary_writer=None,
debug_mode=False):
"""Creates an agent.
Args:
sess: A `tf.compat.v1.Session` object for running associated ops.
environment: A gym environment (e.g. Atari 2600).
agent_name: str, name of the agent to create.
summary_writer: A Tensorflow summary writer to pass to the agent
for in-agent training statistics in Tensorboard.
debug_mode: bool, whether to output Tensorboard summaries. If set to true,
the agent will output in-episode statistics to Tensorboard. Disabled by
default as this results in slower training.
Returns:
agent: An RL agent.
Raises:
ValueError: If `agent_name` is not in supported list.
"""
assert agent_name is not None
if not debug_mode:
summary_writer = None
if agent_name == 'dqn':
return dqn_agent.DQNAgent(sess,
num_actions=environment.action_space.n,
summary_writer=summary_writer)
elif agent_name == 'rainbow':
return rainbow_agent.RainbowAgent(
sess,
num_actions=environment.action_space.n,
summary_writer=summary_writer)
elif agent_name == 'implicit_quantile':
return implicit_quantile_agent.ImplicitQuantileAgent(
sess,
num_actions=environment.action_space.n,
summary_writer=summary_writer)
elif agent_name == 'jax_dqn':
return jax_dqn_agent.JaxDQNAgent(
num_actions=environment.action_space.n,
summary_writer=summary_writer)
elif agent_name == 'jax_quantile':
return jax_quantile_agent.JaxQuantileAgent(
num_actions=environment.action_space.n,
summary_writer=summary_writer)
elif agent_name == 'jax_rainbow':
return jax_rainbow_agent.JaxRainbowAgent(
num_actions=environment.action_space.n,
summary_writer=summary_writer)
elif agent_name == 'jax_implicit_quantile':
return jax_implicit_quantile_agent.JaxImplicitQuantileAgent(
num_actions=environment.action_space.n,
summary_writer=summary_writer)
else:
raise ValueError('Unknown agent: {}'.format(agent_name))
def testCreateAgentWithDefaults(self):
# Verifies that we can create and train an agent with the default values.
with tf.Session() as sess:
agent = rainbow_agent.RainbowAgent(sess, num_actions=4)
sess.run(tf.global_variables_initializer())
observation = np.ones([84, 84, 1])
agent.begin_episode(observation)
agent.step(reward=1, observation=observation)
agent.end_episode(reward=1)
def testStoreTransitionWithPrioritizedSamplingy(self):
with tf.Session() as sess:
agent = rainbow_agent.RainbowAgent(
sess, num_actions=4, replay_scheme='prioritized')
dummy_frame = np.zeros((84, 84))
# Adding transitions with default, 10., default priorities.
agent._store_transition(dummy_frame, 0, 0, False)
agent._store_transition(dummy_frame, 0, 0, False, 10.)
agent._store_transition(dummy_frame, 0, 0, False)
returned_priorities = agent._replay.memory.get_priority(
np.arange(self.stack_size - 1, self.stack_size + 2, dtype=np.int32))
expected_priorities = [1., 10., 10.]
self.assertAllEqual(returned_priorities, expected_priorities)
def __init__(self, **kwargs):
super(RainbowWrapper, self).__init__(**kwargs)
args, num_inputs, num_outputs, factor = self.get_args(kwargs)
self.minmax = default_value_arg(kwargs, 'minmax', None)
if args.true_environment and args.state_forms[0] == 'raw':
network = atari_lib.rainbow_network
print(network)
observation_shape = (84, 84)
else:
network = create_rainbow_network(self.minmax)
print("num inputs", num_inputs)
observation_shape = (num_inputs, )
self.sess = tf.Session(
'', config=tf.ConfigProto(allow_soft_placement=True))
# local_device_protos = device_lib.list_local_devices()
# print([x.name for x in local_device_protos])
# print(atari_lib.NATURE_DQN_OBSERVATION_SHAPE,
# atari_lib.NATURE_DQN_DTYPE,
# atari_lib.NATURE_DQN_STACK_SIZE,)
# print(num_inputs)
self.dope_rainbow = rainbow_agent.RainbowAgent(
self.sess,
num_outputs,
observation_shape=observation_shape,
observation_dtype=tf.float32,
stack_size=args.num_stack,
network=network,
num_atoms=51,
vmax=args.value_bounds[1],
gamma=args.gamma,
update_horizon=1,
min_replay_history=20000,
update_period=4,
target_update_period=8000,
epsilon_fn=dqn_agent.linearly_decaying_epsilon,
epsilon_train=0.01,
epsilon_eval=0.001,
epsilon_decay_period=250000,
replay_scheme='prioritized',
tf_device='/gpu:' + str(args.gpu),
use_staging=True,
optimizer=tf.train.AdamOptimizer(learning_rate=.00025,
epsilon=0.0003125),
summary_writer=None,
summary_writing_frequency=500)
self.dope_rainbow.eval_mode = False
self.sess.run(tf.global_variables_initializer())
def create_agent(sess, agent_name, num_actions,
observation_shape=atari_lib.NATURE_DQN_OBSERVATION_SHAPE,
observation_dtype=atari_lib.NATURE_DQN_DTYPE,
stack_size=atari_lib.NATURE_DQN_STACK_SIZE,
summary_writer=None):
"""Creates an agent.
Args:
sess: A `tf.Session` object for running associated ops.
agent_name: str, name of the agent to create.
num_actions: int, number of actions the agent can take at any state.
summary_writer: A Tensorflow summary writer to pass to the agent
for in-agent training statistics in Tensorboard.
observation_shape: tuple of ints describing the observation shape.
observation_dtype: tf.DType, specifies the type of the observations. Note
that if your inputs are continuous, you should set this to tf.float32.
stack_size: int, number of frames to use in state stack.
Returns:
agent: An RL agent.
Raises:
ValueError: If `agent_name` is not in supported list or one of the
GAIRL submodules is not in supported list when the chosen agent is GAIRL.
"""
if agent_name == 'dqn':
return dqn_agent.DQNAgent(
sess, num_actions, observation_shape=observation_shape,
observation_dtype=observation_dtype, stack_size=stack_size,
summary_writer=summary_writer
)
elif agent_name == 'rainbow':
return rainbow_agent.RainbowAgent(
sess, num_actions, observation_shape=observation_shape,
observation_dtype=observation_dtype, stack_size=stack_size,
summary_writer=summary_writer
)
elif agent_name == 'implicit_quantile':
return implicit_quantile_agent.ImplicitQuantileAgent(
sess, num_actions, summary_writer=summary_writer
)
else:
raise ValueError('Unknown agent: {}'.format(agent_name))
def create_agent(sess, environment):
"""Creates a DQN agent.
Args:
sess: A `tf.Session` object for running associated ops.
environment: An Atari 2600 Gym environment.
Returns:
agent: An RL agent.
Raises:
ValueError: If `agent_name` is not in supported list.
"""
if FLAGS.agent_name == 'dqn':
return dqn_agent.DQNAgent(sess, num_actions=environment.action_space.n)
elif FLAGS.agent_name == 'rainbow':
return rainbow_agent.RainbowAgent(
sess, num_actions=environment.action_space.n)
elif FLAGS.agent_name == 'implicit_quantile':
return implicit_quantile_agent.ImplicitQuantileAgent(
sess, num_actions=environment.action_space.n)
else:
raise ValueError('Unknown agent: {}'.format(FLAGS.agent_name))
def create_agent_fn(sess, env, summary_writer):
return rainbow_agent.RainbowAgent(sess=sess,
num_actions=env.action_space.n,
summary_writer=summary_writer)
def _create_test_agent(self, sess):
stack_size = self.stack_size
# This dummy network allows us to deterministically anticipate that
# action 0 will be selected by an argmax.
# In Rainbow we are dealing with a distribution over Q-values,
# which are represented as num_atoms bins, ranging from -vmax to vmax.
# The output layer will have num_actions * num_atoms elements,
# so each group of num_atoms weights represent the logits for a
# particular action. By setting 1s everywhere, except for the first
# num_atoms (representing the logits for the first action), which are
# set to np.arange(num_atoms), we are ensuring that the first action
# places higher weight on higher Q-values; this results in the first
# action being chosen.
class MockRainbowNetwork(tf.keras.Model):
"""Custom tf.keras.Model used in tests."""
def __init__(self, num_actions, num_atoms, support, **kwargs):
super(MockRainbowNetwork, self).__init__(**kwargs)
self.num_actions = num_actions
self.num_atoms = num_atoms
self.support = support
first_row = np.tile(np.ones(self.num_atoms),
self.num_actions - 1)
first_row = np.concatenate(
(np.arange(self.num_atoms), first_row))
bottom_rows = np.tile(
np.ones(self.num_actions * self.num_atoms),
(stack_size - 1, 1))
weights_initializer = np.concatenate(
([first_row], bottom_rows))
self.layer = tf.keras.layers.Dense(
self.num_actions * self.num_atoms,
kernel_initializer=tf.constant_initializer(
weights_initializer),
bias_initializer=tf.ones_initializer())
def call(self, state):
inputs = tf.constant(np.zeros((state.shape[0], stack_size)),
dtype=tf.float32)
net = self.layer(inputs)
logits = tf.reshape(net,
[-1, self.num_actions, self.num_atoms])
probabilities = tf.keras.activations.softmax(logits)
qs = tf.reduce_sum(self.support * probabilities, axis=2)
return atari_lib.RainbowNetworkType(qs, logits, probabilities)
agent = rainbow_agent.RainbowAgent(
sess=sess,
network=MockRainbowNetwork,
num_actions=self._num_actions,
num_atoms=self._num_atoms,
vmax=self._vmax,
min_replay_history=self._min_replay_history,
epsilon_fn=lambda w, x, y, z: 0.0, # No exploration.
epsilon_eval=0.0,
epsilon_decay_period=self._epsilon_decay_period)
# This ensures non-random action choices (since epsilon_eval = 0.0) and
# skips the train_step.
agent.eval_mode = True
sess.run(tf.global_variables_initializer())
return agent