def testCreateAgentWithPrebuiltPreprocessingLayers(self):
dense_layer = tf.keras.Sequential([
tf.keras.layers.Dense(10),
tf.keras.layers.Flatten(),
tf.keras.layers.Reshape([2, 5]),
])
q_net = KerasLayersNet(self._time_step_spec.observation,
self._action_spec, dense_layer)
with self.assertRaisesRegexp(
ValueError, 'shares weights with the original network'):
categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec,
self._action_spec,
categorical_q_network=q_net,
optimizer=None)
# Explicitly share weights between q and target networks.
# This would be an unusual setup so we check that an error is thrown.
q_target_net = KerasLayersNet(self._time_step_spec.observation,
self._action_spec, dense_layer)
with self.assertRaisesRegexp(
ValueError, 'shares weights with the original network'):
categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec,
self._action_spec,
categorical_q_network=q_net,
optimizer=None,
target_categorical_q_network=q_target_net)
def testCreateAgentWithPrebuiltPreprocessingLayers(self):
dense_layer = tf.keras.layers.Dense(3)
q_net = KerasLayersNet(self._time_step_spec.observation,
self._action_spec, dense_layer)
with self.assertRaisesRegexp(
ValueError, 'shares weights with the original network'):
categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec,
self._action_spec,
categorical_q_network=q_net,
optimizer=None)
# Explicitly share weights between q and target networks; this is ok.
q_target_net = KerasLayersNet(self._time_step_spec.observation,
self._action_spec, dense_layer)
categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec,
self._action_spec,
categorical_q_network=q_net,
optimizer=None,
target_categorical_q_network=q_target_net)
q_bad_target_net = KerasLayersNet(self._time_step_spec.observation,
self._action_spec,
dense_layer,
num_atoms=3)
with self.assertRaisesRegexp(ValueError,
'have different numbers of atoms'):
categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec,
self._action_spec,
categorical_q_network=q_net,
optimizer=None,
target_categorical_q_network=q_bad_target_net)
def testCreateAgentDimChecks(self):
action_spec = tensor_spec.BoundedTensorSpec([1, 2], tf.int32, 0, 1)
with self.assertRaisesRegex(ValueError, 'Only scalar actions'):
categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec, action_spec, self._dummy_categorical_net,
self._optimizer)
def testInitialize(self):
agent = categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec,
self._action_spec,
self._categorical_net,
self._optimizer)
observations = tf.constant([[1, 2], [3, 4]], dtype=tf.float32)
time_steps = ts.restart(observations, batch_size=2)
actions = tf.constant([0, 1], dtype=tf.int32)
action_steps = policy_step.PolicyStep(actions)
rewards = tf.constant([10, 20], dtype=tf.float32)
discounts = tf.constant([0.9, 0.9], dtype=tf.float32)
next_time_steps = ts.transition(observations, rewards, discounts)
experience = test_utils.stacked_trajectory_from_transition(
time_steps, action_steps, next_time_steps)
loss_info = agent._loss(experience)
initialize = agent.initialize()
self.evaluate(tf.compat.v1.global_variables_initializer())
losses = self.evaluate(loss_info).loss
self.assertGreater(losses, 0.0)
critic_variables = agent._q_network.variables
target_critic_variables = agent._target_q_network.variables
self.assertTrue(critic_variables)
self.assertTrue(target_critic_variables)
self.evaluate(initialize)
for s, t in zip(critic_variables, target_critic_variables):
self.assertAllClose(self.evaluate(s), self.evaluate(t))
def testTrain(self):
agent = categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec,
self._action_spec,
self._dummy_categorical_net,
self._optimizer)
observations = tf.constant([[1, 2], [3, 4]], dtype=tf.float32)
time_steps = ts.restart(observations, batch_size=2)
actions = tf.constant([0, 1], dtype=tf.int32)
action_steps = policy_step.PolicyStep(actions)
rewards = tf.constant([10, 20], dtype=tf.float32)
discounts = tf.constant([0.9, 0.9], dtype=tf.float32)
next_observations = tf.constant([[5, 6], [7, 8]], dtype=tf.float32)
next_time_steps = ts.transition(next_observations, rewards, discounts)
experience = test_utils.stacked_trajectory_from_transition(
time_steps, action_steps, next_time_steps)
train_step = agent.train(experience, weights=None)
# Due to the constant initialization of the DummyCategoricalNet, we can
# expect the same loss every time.
expected_loss = 2.19525
self.evaluate(tf.compat.v1.global_variables_initializer())
evaluated_loss, _ = self.evaluate(train_step)
self.assertAllClose(evaluated_loss, expected_loss, atol=1e-4)
def testCriticLossNStep(self):
agent = categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec,
self._action_spec,
self._dummy_categorical_net,
self._optimizer,
n_step_update=2)
observations = tf.constant([[1, 2], [3, 4]], dtype=tf.float32)
time_steps = ts.restart(observations, batch_size=2)
actions = tf.constant([0, 1], dtype=tf.int32)
action_steps = policy_step.PolicyStep(actions)
rewards = tf.constant([10, 20], dtype=tf.float32)
discounts = tf.constant([0.9, 0.9], dtype=tf.float32)
next_observations = tf.constant([[5, 6], [7, 8]], dtype=tf.float32)
next_time_steps = ts.transition(next_observations, rewards, discounts)
third_observations = tf.constant([[9, 10], [11, 12]], dtype=tf.float32)
third_time_steps = ts.transition(third_observations, rewards,
discounts)
experience1 = trajectory.from_transition(time_steps, action_steps,
next_time_steps)
experience2 = trajectory.from_transition(next_time_steps, action_steps,
third_time_steps)
experience3 = trajectory.from_transition(third_time_steps,
action_steps,
third_time_steps)
experience = tf.nest.map_structure(
lambda x, y, z: tf.stack([x, y, z], axis=1), experience1,
experience2, experience3)
loss_info = agent._loss(experience)
# discounted_returns should evaluate to 10 + 0.9 * 10 = 19 and
# 20 + 0.9 * 20 = 38.
evaluated_discounted_returns = self.evaluate(agent._discounted_returns)
self.assertAllClose(evaluated_discounted_returns, [[19], [38]],
atol=1e-4)
# Both final_value_discount values should be 0.9 * 0.9 = 0.81.
evaluated_final_value_discount = self.evaluate(
agent._final_value_discount)
self.assertAllClose(evaluated_final_value_discount, [[0.81], [0.81]],
atol=1e-4)
# Due to the constant initialization of the DummyCategoricalNet, we can
# expect the same loss every time.
expected_loss = 2.19525
self.evaluate(tf.compat.v1.global_variables_initializer())
evaluated_loss = self.evaluate(loss_info).loss
self.assertAllClose(evaluated_loss, expected_loss, atol=1e-4)
def testCreateAgentNestSizeChecks(self):
action_spec = [
tensor_spec.BoundedTensorSpec([1], tf.int32, 0, 1),
tensor_spec.BoundedTensorSpec([1], tf.int32, 0, 1)
]
with self.assertRaisesRegexp(ValueError,
'.*Only one dimensional actions.*'):
categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec, action_spec, self._dummy_categorical_net,
self._optimizer)
def testTrainWithRnn(self):
action_spec = tensor_spec.BoundedTensorSpec([1], tf.int32, 0, 1)
batch_size = 5
observations = tf.constant([[[1, 2], [3, 4], [5, 6]]] * batch_size,
dtype=tf.float32)
actions = tf.constant([[[0], [1], [1]]] * batch_size, dtype=tf.int32)
time_steps = ts.TimeStep(step_type=tf.constant([[1] * 3] * batch_size,
dtype=tf.int32),
reward=tf.constant([[1] * 3] * batch_size,
dtype=tf.float32),
discount=tf.constant([[1] * 3] * batch_size,
dtype=tf.float32),
observation=[observations])
experience = trajectory.Trajectory(step_type=time_steps.step_type,
observation=observations,
action=actions,
policy_info=(),
next_step_type=time_steps.step_type,
reward=time_steps.reward,
discount=time_steps.discount)
categorical_q_rnn_network = DummyCategoricalQRnnNetwork(
self._obs_spec,
action_spec,
conv_layer_params=None,
input_fc_layer_params=(16, ),
preprocessing_combiner=None,
lstm_size=(40, ),
output_fc_layer_params=(16, ),
)
counter = common.create_variable('test_train_counter')
agent = categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec,
action_spec,
categorical_q_rnn_network,
optimizer=tf.compat.v1.train.AdamOptimizer(0.001),
)
# Force variable creation.
agent.policy.variables()
if tf.executing_eagerly():
loss = lambda: agent.train(experience)
else:
loss = agent.train(experience)
self.evaluate(tf.compat.v1.global_variables_initializer())
self.assertEqual(self.evaluate(counter), 0)
self.evaluate(loss)
def testPolicy(self):
agent = categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec, self._action_spec, self._categorical_net,
self._optimizer)
observations = tf.constant([[1, 2], [3, 4]], dtype=tf.float32)
time_steps = ts.restart(observations, batch_size=2)
actions, _, _ = agent.policy.action(time_steps)
self.assertEqual(actions.shape, [2])
self.evaluate(tf.compat.v1.global_variables_initializer())
actions_ = self.evaluate(actions)
self.assertTrue(all(actions_ <= self._action_spec.maximum))
self.assertTrue(all(actions_ >= self._action_spec.minimum))
def testCreateAgentWithPrebuiltPreprocessingLayersDiffAtoms(self):
dense_layer = tf.keras.layers.Dense(3)
q_net = KerasLayersNet(
self._time_step_spec.observation, self._action_spec, dense_layer)
dense_layer_target = tf.keras.layers.Dense(3)
q_bad_target_net = KerasLayersNet(
self._time_step_spec.observation, self._action_spec, dense_layer_target,
num_atoms=3)
with self.assertRaisesRegexp(ValueError, 'have different numbers of atoms'):
categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec,
self._action_spec,
categorical_q_network=q_net,
optimizer=None,
target_categorical_q_network=q_bad_target_net)
def testCriticLossWithMaskedActions(self):
# Observations are now a tuple of the usual observation and an action mask.
observation_spec_with_mask = (self._obs_spec,
tensor_spec.BoundedTensorSpec([2],
tf.int32,
0, 1))
time_step_spec = ts.time_step_spec(observation_spec_with_mask)
dummy_categorical_net = DummyCategoricalNet(self._obs_spec)
agent = categorical_dqn_agent.CategoricalDqnAgent(
time_step_spec,
self._action_spec,
dummy_categorical_net,
self._optimizer,
observation_and_action_constraint_splitter=lambda x: (x[0], x[1]))
# For `observations`, the masks are set up so that only one action is valid
# for each element in the batch.
observations = (tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
tf.constant([[1, 0], [0, 1]], dtype=tf.int32))
time_steps = ts.restart(observations, batch_size=2)
actions = tf.constant([0, 1], dtype=tf.int32)
action_steps = policy_step.PolicyStep(actions)
rewards = tf.constant([10, 20], dtype=tf.float32)
discounts = tf.constant([0.9, 0.9], dtype=tf.float32)
# For `next_observations`, the masks are set up so the opposite actions as
# before are valid.
next_observations = (tf.constant([[5, 6], [7, 8]], dtype=tf.float32),
tf.constant([[0, 1], [1, 0]], dtype=tf.int32))
next_time_steps = ts.transition(next_observations, rewards, discounts)
experience = test_utils.stacked_trajectory_from_transition(
time_steps, action_steps, next_time_steps)
# Due to the constant initialization of the DummyCategoricalNet, we can
# expect the same loss every time. Note this is different from the loss in
# testCriticLoss above due to previously optimal actions being masked out.
expected_loss = 5.062895
loss_info = agent._loss(experience)
self.evaluate(tf.compat.v1.global_variables_initializer())
evaluated_loss = self.evaluate(loss_info).loss
self.assertAllClose(evaluated_loss, expected_loss, atol=1e-4)
def testUpdateTarget(self):
agent = categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec, self._action_spec, self._categorical_net,
self._optimizer)
observations = tf.constant([[1, 2], [3, 4]], dtype=tf.float32)
time_steps = ts.restart(observations, batch_size=2)
actions = tf.constant([0, 1], dtype=tf.int32)
action_steps = policy_step.PolicyStep(actions)
experience = test_utils.stacked_trajectory_from_transition(
time_steps, action_steps, time_steps)
loss_info = agent._loss(experience)
update_targets = agent._update_target()
self.evaluate(tf.compat.v1.global_variables_initializer())
losses = self.evaluate(loss_info).loss
self.assertGreater(losses, 0.0)
self.evaluate(update_targets)
def load_agents_and_create_videos(
root_dir,
env_name='CartPole-v0',
num_iterations=NUM_ITERATIONS,
max_ep_steps=1000,
train_sequence_length=1,
# Params for QNetwork
fc_layer_params=((100, )),
# Params for QRnnNetwork
input_fc_layer_params=(50, ),
lstm_size=(20, ),
output_fc_layer_params=(20, ),
# Params for collect
initial_collect_steps=10000,
collect_steps_per_iteration=1,
epsilon_greedy=0.1,
replay_buffer_capacity=100000,
# Params for target update
target_update_tau=0.05,
target_update_period=5,
# Params for train
train_steps_per_iteration=1,
batch_size=64,
learning_rate=1e-3,
num_atoms=51,
min_q_value=-20,
max_q_value=20,
n_step_update=1,
gamma=0.99,
reward_scale_factor=1.0,
gradient_clipping=None,
use_tf_functions=True,
# Params for eval
num_eval_episodes=10,
num_random_episodes=1,
eval_interval=1000,
# Params for checkpoints
train_checkpoint_interval=10000,
policy_checkpoint_interval=5000,
rb_checkpoint_interval=20000,
# Params for summaries and logging
log_interval=1000,
summary_interval=1000,
summaries_flush_secs=10,
debug_summaries=False,
summarize_grads_and_vars=False,
eval_metrics_callback=None,
random_metrics_callback=None):
# Define the directories to read from
train_dir = os.path.join(root_dir, 'train')
eval_dir = os.path.join(root_dir, 'eval')
random_dir = os.path.join(root_dir, 'random')
# Match the writers and metrics used in training
train_summary_writer = tf.compat.v2.summary.create_file_writer(
train_dir, flush_millis=summaries_flush_secs * 1000)
train_summary_writer.set_as_default()
eval_summary_writer = tf.compat.v2.summary.create_file_writer(
eval_dir, flush_millis=summaries_flush_secs * 1000)
eval_metrics = [
tf_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
tf_metrics.AverageEpisodeLengthMetric(buffer_size=num_eval_episodes)
]
random_summary_writer = tf.compat.v2.summary.create_file_writer(
random_dir, flush_millis=summaries_flush_secs * 1000)
random_metrics = [
tf_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
tf_metrics.AverageEpisodeLengthMetric(buffer_size=num_eval_episodes)
]
global_step = tf.compat.v1.train.get_or_create_global_step()
# Match the environments used in training
tf_env = tf_py_environment.TFPyEnvironment(
suite_gym.load(env_name, max_episode_steps=max_ep_steps))
eval_py_env = suite_gym.load(env_name, max_episode_steps=max_ep_steps)
eval_tf_env = tf_py_environment.TFPyEnvironment(eval_py_env)
# Match the agents used in training
categorical_q_net = categorical_q_network.CategoricalQNetwork(
tf_env.observation_spec(),
tf_env.action_spec(),
num_atoms=num_atoms,
fc_layer_params=fc_layer_params)
optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=learning_rate)
tf_agent = categorical_dqn_agent.CategoricalDqnAgent(
tf_env.time_step_spec(),
tf_env.action_spec(),
categorical_q_network=categorical_q_net,
optimizer=optimizer,
min_q_value=min_q_value,
max_q_value=max_q_value,
n_step_update=n_step_update,
td_errors_loss_fn=common.element_wise_squared_loss,
gamma=gamma,
train_step_counter=global_step)
tf_agent.initialize()
train_metrics = [
# tf_metrics.NumberOfEpisodes(),
# tf_metrics.EnvironmentSteps(),
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
eval_policy = tf_agent.policy
collect_policy = tf_agent.collect_policy
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
data_spec=tf_agent.collect_data_spec,
batch_size=tf_env.batch_size,
max_length=replay_buffer_capacity)
collect_driver = dynamic_step_driver.DynamicStepDriver(
tf_env,
collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_steps=collect_steps_per_iteration)
train_checkpointer = common.Checkpointer(ckpt_dir=train_dir,
agent=tf_agent,
global_step=global_step,
metrics=metric_utils.MetricsGroup(
train_metrics,
'train_metrics'))
policy_checkpointer = common.Checkpointer(ckpt_dir=os.path.join(
train_dir, 'policy'),
policy=eval_policy,
global_step=global_step)
rb_checkpointer = common.Checkpointer(ckpt_dir=os.path.join(
train_dir, 'replay_buffer'),
max_to_keep=1,
replay_buffer=replay_buffer)
train_checkpointer.initialize_or_restore()
rb_checkpointer.initialize_or_restore()
if use_tf_functions:
# To speed up collect use common.function.
collect_driver.run = common.function(collect_driver.run)
tf_agent.train = common.function(tf_agent.train)
random_policy = random_tf_policy.RandomTFPolicy(
eval_tf_env.time_step_spec(), eval_tf_env.action_spec())
# Make movies of the trained agent and a random agent
date_string = datetime.datetime.now().strftime('%Y-%m-%d_%H%M%S')
# Finally, used the saved policy to generate the video
trained_filename = "trainedC51_" + date_string
create_policy_eval_video(eval_tf_env, eval_py_env, tf_agent.policy,
trained_filename)
# And, create one with a random agent for comparison
random_filename = 'random_' + date_string
create_policy_eval_video(eval_tf_env, eval_py_env, random_policy,
random_filename)
def main(argv):
tf.compat.v1.enable_v2_behavior()
logging.config.dictConfig({
'version': 1,
# Other configs ...
'disable_existing_loggers': True
})
argv = argv[0]
evaluate = argv.eval
# Mostly copied from https://www.tensorflow.org/agents/tutorials/1_dqn_tutorial
# Hyperparameters
num_iterations = argv.num_iterations
collect_steps_per_iteration = argv.collect_steps_per_iteration
replay_buffer_max_length = 100000
batch_size = argv.batch_size
learning_rate = 2.5e-5
log_interval = argv.log_interval
num_atoms = argv.num_atoms
min_q_value = argv.min_q_value
max_q_value = argv.max_q_value
n_step_update = argv.n_step_update
gamma = 0.99
num_eval_episodes = 10
eval_interval = argv.eval_interval
save_interval = argv.save_interval
n_parallels = argv.n_parallels
train_in_browser = argv.train_in_browser
# Environment
train_py_env = Env2048(evaluate) if evaluate else ParallelPyEnvironment(
[lambda: Env2048(train_in_browser)] * n_parallels,
start_serially=False)
eval_py_env = Env2048(evaluate)
train_env = tf_py_environment.TFPyEnvironment(train_py_env)
eval_env = tf_py_environment.TFPyEnvironment(eval_py_env)
# Agent
fc_layer_params = (64, 64, 32)
conv_layer_params = ((512, (2, 1), (1, 1)), (512, (1, 2), (1, 1)))
preprocessing_layers = tf.keras.models.Sequential([
tf.keras.layers.Conv2D(512, (1, 1), (1, 1), padding='same'),
tf.keras.layers.Conv2D(512, (2, 1), (1, 1), padding='same'),
tf.keras.layers.Conv2D(512, (1, 2), (1, 1), padding='same'),
tf.keras.layers.Flatten()
])
preprocessing_combiner = tf.keras.layers.Concatenate(axis=-1)
optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)
global_step = tf.compat.v1.train.get_or_create_global_step()
# q_net = q_network.QNetwork(
# train_env.observation_spec(),
# train_env.action_spec(),
# fc_layer_params=fc_layer_params)
# agent = dqn_agent.DqnAgent(
# train_env.time_step_spec(),
# train_env.action_spec(),
# q_network=q_net,
# optimizer=optimizer,
# td_errors_loss_fn=common.element_wise_squared_loss,
# train_step_counter=global_step)
categorical_q_net = categorical_q_network.CategoricalQNetwork(
train_env.observation_spec(),
train_env.action_spec(),
num_atoms=num_atoms,
fc_layer_params=fc_layer_params,
# conv_layer_params=conv_layer_params
preprocessing_layers=preprocessing_layers,
preprocessing_combiner=preprocessing_combiner)
agent = categorical_dqn_agent.CategoricalDqnAgent(
train_env.time_step_spec(),
train_env.action_spec(),
categorical_q_network=categorical_q_net,
optimizer=optimizer,
min_q_value=min_q_value,
max_q_value=max_q_value,
n_step_update=n_step_update,
td_errors_loss_fn=common.element_wise_squared_loss,
gamma=gamma,
train_step_counter=global_step)
agent.initialize()
# Replay buffer
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
data_spec=agent.collect_data_spec,
batch_size=train_env.batch_size,
max_length=replay_buffer_max_length)
# Data Collection
collect_driver = dynamic_step_driver.DynamicStepDriver(
train_env,
agent.collect_policy,
observers=[replay_buffer.add_batch],
num_steps=collect_steps_per_iteration)
collect_driver.run()
dataset = replay_buffer.as_dataset(num_parallel_calls=3,
sample_batch_size=batch_size,
num_steps=2).prefetch(3)
iterator = iter(dataset)
# Checkpointer
checkpoint_dir = os.path.join(os.getcwd(), 'checkpoint')
train_checkpointer = common.Checkpointer(ckpt_dir=checkpoint_dir,
max_to_keep=1,
agent=agent,
policy=agent.policy,
replay_buffer=replay_buffer,
global_step=global_step)
train_checkpointer.initialize_or_restore()
global_step = tf.compat.v1.train.get_global_step()
# Training
if evaluate:
avg_return, best_eval_score = compute_avg_return(
eval_env, agent.policy, num_eval_episodes)
print(f"Average return: {avg_return}, best score = {best_eval_score}")
train_env.station.shutdown()
eval_env.station.shutdown()
else:
agent.train = common.function(agent.train)
# agent.train_step_counter.assign(0)
avg_return = compute_avg_return(eval_env, agent.policy,
num_eval_episodes)
returns = [avg_return]
t = trange(global_step.numpy(), num_iterations, leave=True)
best_scores = np.array(
list(map(lambda env: env.best_score, train_env.envs)))
for _ in t:
# Collect a few steps using collect_policy and save to the replay buffer.
collect_driver.run()
# Sample a batch of data from the buffer and update the agent's network.
experience, unused_info = next(iterator)
train_loss = agent.train(experience).loss
scores = list(map(lambda env: env.score, train_env.envs))
t.set_description(desc=f"Scores = {scores}")
step = tf.compat.v1.train.get_global_step().numpy()
if step % log_interval == 0:
t.write(f"step = {step}: loss = {train_loss}")
if step % save_interval == 0:
train_checkpointer.save(step)
if step % eval_interval == 0:
avg_return, best_eval_score = compute_avg_return(
eval_env, agent.policy, num_eval_episodes)
new_best_scores = np.array(
list(map(lambda env: env.best_score, train_env.envs)))
diff = np.subtract(new_best_scores, best_scores)
best_scores = new_best_scores
if np.count_nonzero(diff) > 0:
t.write(f"step = {step}: Best scores = {best_scores}")
t.write(
f'step = {step}: Average Return = {avg_return}, best score reached in training = '
f'{max(list(map(lambda env: env.best_score, train_env.envs)))}'
f', best score in eval = {best_eval_score}')
returns.append(avg_return)
steps = range(0, num_iterations + 1, eval_interval)
plt.plot(steps, returns)
plt.ylabel('Average Return')
plt.xlabel('Step')
train_env.close()
eval_env.close()
train_py_env.close()
def testCreateAgentDefaultNetwork(self):
categorical_dqn_agent.CategoricalDqnAgent(
self._time_step_spec,
self._action_spec,
self._categorical_net,
self._optimizer)
momentum=optimizer_momentum,
epsilon=optimizer_epsilon,
centered=True)
# Computes epsilon for epsilon greedy policy given the training step
epsilon_fn = keras.optimizers.schedules.PolynomialDecay(
initial_learning_rate=1.0, # initial ε
decay_steps=epsilon_decay_steps,
end_learning_rate=epsilon_final) # final ε
agent = categorical_dqn_agent.CategoricalDqnAgent(
tf_env.time_step_spec(),
tf_env.action_spec(),
categorical_q_network=categorical_q_net,
optimizer=optimizer,
min_q_value=int(min_q_value),
max_q_value=int(max_q_value),
n_step_update=int(n_step_update),
td_errors_loss_fn=common.element_wise_squared_loss,
gamma=discount_factor,
train_step_counter=train_step,
epsilon_greedy=lambda: epsilon_fn(train_step))
agent.initialize()
# Speed up as tensorflow function
agent.train = function(agent.train)
## ------------------------------------------------------------------------------
## ------------------------------------------------------------------------------
## ------------------------------------------------------------------------------
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
def __init__(
self,
root_dir,
env_name,
num_iterations=200,
max_episode_frames=108000, # ALE frames
terminal_on_life_loss=False,
conv_layer_params=((32, (8, 8), 4), (64, (4, 4), 2), (64, (3, 3),
1)),
fc_layer_params=(512, ),
# Params for collect
initial_collect_steps=80000, # ALE frames
epsilon_greedy=0.01,
epsilon_decay_period=1000000, # ALE frames
replay_buffer_capacity=1000000,
# Params for train
train_steps_per_iteration=1000000, # ALE frames
update_period=16, # ALE frames
target_update_tau=1.0,
target_update_period=32000, # ALE frames
batch_size=32,
learning_rate=2.5e-4,
n_step_update=2,
gamma=0.99,
reward_scale_factor=1.0,
gradient_clipping=None,
# Params for eval
do_eval=True,
eval_steps_per_iteration=500000, # ALE frames
eval_epsilon_greedy=0.001,
# Params for checkpoints, summaries, and logging
log_interval=1000,
summary_interval=1000,
summaries_flush_secs=10,
debug_summaries=True,
summarize_grads_and_vars=True,
eval_metrics_callback=None):
"""A simple Atari train and eval for DQN.
Args:
root_dir: Directory to write log files to.
env_name: Fully-qualified name of the Atari environment (i.e. Pong-v0).
num_iterations: Number of train/eval iterations to run.
max_episode_frames: Maximum length of a single episode, in ALE frames.
terminal_on_life_loss: Whether to simulate an episode termination when a
life is lost.
conv_layer_params: Params for convolutional layers of QNetwork.
fc_layer_params: Params for fully connected layers of QNetwork.
initial_collect_steps: Number of frames to ALE frames to process before
beginning to train. Since this is in ALE frames, there will be
initial_collect_steps/4 items in the replay buffer when training starts.
epsilon_greedy: Final epsilon value to decay to for training.
epsilon_decay_period: Period over which to decay epsilon, from 1.0 to
epsilon_greedy (defined above).
replay_buffer_capacity: Maximum number of items to store in the replay
buffer.
train_steps_per_iteration: Number of ALE frames to run through for each
iteration of training.
update_period: Run a train operation every update_period ALE frames.
target_update_tau: Coeffecient for soft target network updates (1.0 ==
hard updates).
target_update_period: Period, in ALE frames, to copy the live network to
the target network.
batch_size: Number of frames to include in each training batch.
learning_rate: RMS optimizer learning rate.
n_step_update: The number of steps to consider when computing TD error and
TD loss. Applies standard single-step updates when set to 1.
gamma: Discount for future rewards.
reward_scale_factor: Scaling factor for rewards.
gradient_clipping: Norm length to clip gradients.
do_eval: If True, run an eval every iteration. If False, skip eval.
eval_steps_per_iteration: Number of ALE frames to run through for each
iteration of evaluation.
eval_epsilon_greedy: Epsilon value to use for the evaluation policy (0 ==
totally greedy policy).
log_interval: Log stats to the terminal every log_interval training
steps.
summary_interval: Write TF summaries every summary_interval training
steps.
summaries_flush_secs: Flush summaries to disk every summaries_flush_secs
seconds.
debug_summaries: If True, write additional summaries for debugging (see
dqn_agent for which summaries are written).
summarize_grads_and_vars: Include gradients in summaries.
eval_metrics_callback: A callback function that takes (metric_dict,
global_step) as parameters. Called after every eval with the results of
the evaluation.
"""
self._update_period = update_period / ATARI_FRAME_SKIP
self._train_steps_per_iteration = (train_steps_per_iteration /
ATARI_FRAME_SKIP)
self._do_eval = do_eval
self._eval_steps_per_iteration = eval_steps_per_iteration / ATARI_FRAME_SKIP
self._eval_epsilon_greedy = eval_epsilon_greedy
self._initial_collect_steps = initial_collect_steps / ATARI_FRAME_SKIP
self._summary_interval = summary_interval
self._num_iterations = num_iterations
self._log_interval = log_interval
self._eval_metrics_callback = eval_metrics_callback
with gin.unlock_config():
gin.bind_parameter(('tf_agents.environments.atari_preprocessing.'
'AtariPreprocessing.terminal_on_life_loss'),
terminal_on_life_loss)
root_dir = os.path.expanduser(root_dir)
train_dir = os.path.join(root_dir, 'train')
eval_dir = os.path.join(root_dir, 'eval')
train_summary_writer = tf.compat.v2.summary.create_file_writer(
train_dir, flush_millis=summaries_flush_secs * 1000)
train_summary_writer.set_as_default()
self._train_summary_writer = train_summary_writer
self._eval_summary_writer = None
if self._do_eval:
self._eval_summary_writer = tf.compat.v2.summary.create_file_writer(
eval_dir, flush_millis=summaries_flush_secs * 1000)
self._eval_metrics = [
py_metrics.AverageReturnMetric(name='PhaseAverageReturn',
buffer_size=np.inf),
py_metrics.AverageEpisodeLengthMetric(
name='PhaseAverageEpisodeLength', buffer_size=np.inf),
]
self._global_step = tf.compat.v1.train.get_or_create_global_step()
with tf.compat.v2.summary.record_if(lambda: tf.math.equal(
self._global_step % self._summary_interval, 0)):
self._env = suite_atari.load(
env_name,
max_episode_steps=max_episode_frames / ATARI_FRAME_SKIP,
gym_env_wrappers=suite_atari.
DEFAULT_ATARI_GYM_WRAPPERS_WITH_STACKING)
self._env = batched_py_environment.BatchedPyEnvironment(
[self._env])
observation_spec = tensor_spec.from_spec(
self._env.observation_spec())
time_step_spec = ts.time_step_spec(observation_spec)
action_spec = tensor_spec.from_spec(self._env.action_spec())
with tf.device('/cpu:0'):
epsilon = tf.compat.v1.train.polynomial_decay(
1.0,
self._global_step,
epsilon_decay_period / ATARI_FRAME_SKIP /
self._update_period,
end_learning_rate=epsilon_greedy)
with tf.device('/gpu:0'):
optimizer = tf.compat.v1.train.RMSPropOptimizer(
learning_rate=learning_rate,
decay=0.95,
momentum=0.0,
epsilon=0.00001,
centered=True)
categorical_q_net = AtariCategoricalQNetwork(
observation_spec,
action_spec,
conv_layer_params=conv_layer_params,
fc_layer_params=fc_layer_params)
agent = categorical_dqn_agent.CategoricalDqnAgent(
time_step_spec,
action_spec,
categorical_q_network=categorical_q_net,
optimizer=optimizer,
epsilon_greedy=epsilon,
n_step_update=n_step_update,
target_update_tau=target_update_tau,
target_update_period=(target_update_period /
ATARI_FRAME_SKIP /
self._update_period),
gamma=gamma,
reward_scale_factor=reward_scale_factor,
gradient_clipping=gradient_clipping,
debug_summaries=debug_summaries,
summarize_grads_and_vars=summarize_grads_and_vars,
train_step_counter=self._global_step)
self._collect_policy = py_tf_policy.PyTFPolicy(
agent.collect_policy)
if self._do_eval:
self._eval_policy = py_tf_policy.PyTFPolicy(
epsilon_greedy_policy.EpsilonGreedyPolicy(
policy=agent.policy,
epsilon=self._eval_epsilon_greedy))
py_observation_spec = self._env.observation_spec()
py_time_step_spec = ts.time_step_spec(py_observation_spec)
py_action_spec = policy_step.PolicyStep(
self._env.action_spec())
data_spec = trajectory.from_transition(py_time_step_spec,
py_action_spec,
py_time_step_spec)
self._replay_buffer = py_hashed_replay_buffer.PyHashedReplayBuffer(
data_spec=data_spec, capacity=replay_buffer_capacity)
with tf.device('/cpu:0'):
ds = self._replay_buffer.as_dataset(
sample_batch_size=batch_size, num_steps=n_step_update + 1)
ds = ds.prefetch(4)
ds = ds.apply(
tf.data.experimental.prefetch_to_device('/gpu:0'))
with tf.device('/gpu:0'):
self._ds_itr = tf.compat.v1.data.make_one_shot_iterator(ds)
experience = self._ds_itr.get_next()
self._train_op = agent.train(experience)
self._env_steps_metric = py_metrics.EnvironmentSteps()
self._step_metrics = [
py_metrics.NumberOfEpisodes(),
self._env_steps_metric,
]
self._train_metrics = self._step_metrics + [
py_metrics.AverageReturnMetric(buffer_size=10),
py_metrics.AverageEpisodeLengthMetric(buffer_size=10),
]
# The _train_phase_metrics average over an entire train iteration,
# rather than the rolling average of the last 10 episodes.
self._train_phase_metrics = [
py_metrics.AverageReturnMetric(name='PhaseAverageReturn',
buffer_size=np.inf),
py_metrics.AverageEpisodeLengthMetric(
name='PhaseAverageEpisodeLength', buffer_size=np.inf),
]
self._iteration_metric = py_metrics.CounterMetric(
name='Iteration')
# Summaries written from python should run every time they are
# generated.
with tf.compat.v2.summary.record_if(True):
self._steps_per_second_ph = tf.compat.v1.placeholder(
tf.float32, shape=(), name='steps_per_sec_ph')
self._steps_per_second_summary = tf.compat.v2.summary.scalar(
name='global_steps_per_sec',
data=self._steps_per_second_ph,
step=self._global_step)
for metric in self._train_metrics:
metric.tf_summaries(train_step=self._global_step,
step_metrics=self._step_metrics)
for metric in self._train_phase_metrics:
metric.tf_summaries(
train_step=self._global_step,
step_metrics=(self._iteration_metric, ))
self._iteration_metric.tf_summaries(
train_step=self._global_step)
if self._do_eval:
with self._eval_summary_writer.as_default():
for metric in self._eval_metrics:
metric.tf_summaries(
train_step=self._global_step,
step_metrics=(self._iteration_metric, ))
self._train_checkpointer = common.Checkpointer(
ckpt_dir=train_dir,
agent=agent,
global_step=self._global_step,
optimizer=optimizer,
metrics=metric_utils.MetricsGroup(
self._train_metrics + self._train_phase_metrics +
[self._iteration_metric], 'train_metrics'))
self._policy_checkpointer = common.Checkpointer(
ckpt_dir=os.path.join(train_dir, 'policy'),
policy=agent.policy,
global_step=self._global_step)
self._rb_checkpointer = common.Checkpointer(
ckpt_dir=os.path.join(train_dir, 'replay_buffer'),
max_to_keep=1,
replay_buffer=self._replay_buffer)
self._init_agent_op = agent.initialize()
def create_agent(
agent_class,
environment,
fc_layer_params,
learning_rate,
decaying_epsilon,
n_step_update,
target_update_tau,
target_update_period,
gamma,
reward_scale_factor,
gradient_clipping,
debug_summaries,
summarize_grads_and_vars,
train_step_counter,
num_atoms=None, # Only for categorical_dqn
min_q_value=None, # Only for categorical_dqn
max_q_value=None, # Only for categorical_dqn
):
"""Creates the Hanabi agent.
Args:
agent_class: str, type of agent to construct.
environment: The environment.
learning_rate: The Learning Rate
decaying_epsilon: Epsilon for Epsilon Greedy Policy
target_update_tau: Agent parameter
target_update_period: Agent parameter
gamma: Agent parameter
reward_scale_factor: Agent parameter
gradient_clipping: Agent parameter
debug_summaries: Agent parameter
summarize_grads_and_vars: Agent parameter
train_step_counter: The train step tf.Variable to be passed to agent
Returns:
An agent for playing Hanabi.
Raises:
ValueError: if an unknown agent type is requested.
"""
if agent_class == 'DQN':
return dqn_agent.DqnAgent(
environment.time_step_spec(),
environment.action_spec(),
q_network=q_network.QNetwork(
environment.time_step_spec().observation['observations'],
environment.action_spec(),
fc_layer_params=fc_layer_params),
optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
observation_and_action_constraint_splitter=
observation_and_action_constraint_splitter,
epsilon_greedy=decaying_epsilon,
n_step_update=n_step_update,
target_update_tau=target_update_tau,
target_update_period=target_update_period,
td_errors_loss_fn=common.element_wise_squared_loss,
gamma=gamma,
reward_scale_factor=reward_scale_factor,
gradient_clipping=gradient_clipping,
debug_summaries=debug_summaries,
summarize_grads_and_vars=summarize_grads_and_vars,
train_step_counter=train_step_counter)
elif agent_class == 'DDQN':
return dqn_agent.DdqnAgent(
environment.time_step_spec(),
environment.action_spec(),
q_network=q_network.QNetwork(
environment.time_step_spec().observation['observations'],
environment.action_spec(),
fc_layer_params=fc_layer_params),
optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
observation_and_action_constraint_splitter=
observation_and_action_constraint_splitter,
epsilon_greedy=decaying_epsilon,
n_step_update=n_step_update,
target_update_tau=target_update_tau,
target_update_period=target_update_period,
td_errors_loss_fn=common.element_wise_squared_loss,
gamma=gamma,
reward_scale_factor=reward_scale_factor,
gradient_clipping=gradient_clipping,
debug_summaries=debug_summaries,
summarize_grads_and_vars=summarize_grads_and_vars,
train_step_counter=train_step_counter)
elif agent_class == 'categorical_dqn':
return categorical_dqn_agent.CategoricalDqnAgent(
environment.time_step_spec(),
environment.action_spec(),
categorical_q_network=categorical_q_network.CategoricalQNetwork(
environment.time_step_spec().observation['observations'],
environment.action_spec(),
num_atoms=num_atoms,
fc_layer_params=fc_layer_params),
optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
observation_and_action_constraint_splitter=
observation_and_action_constraint_splitter,
epsilon_greedy=decaying_epsilon,
n_step_update=n_step_update,
target_update_tau=target_update_tau,
target_update_period=target_update_period,
min_q_value=min_q_value,
max_q_value=max_q_value,
td_errors_loss_fn=common.element_wise_squared_loss,
gamma=gamma,
reward_scale_factor=reward_scale_factor,
gradient_clipping=gradient_clipping,
debug_summaries=debug_summaries,
summarize_grads_and_vars=summarize_grads_and_vars,
train_step_counter=train_step_counter)
else:
raise ValueError(
'Expected valid agent_type, got {}'.format(agent_class))
def main(arg, pars):
"""
"""
print("load env ..")
env_name = ("Car-v0")
#env = gym.make("Car-v0")
env = suite_gym.load(env_name,
discount=arg.gamma,
max_episode_steps=arg.max_t)
print_parameter(arg, pars)
train_py_env = suite_gym.load(env_name,
discount=arg.gamma,
max_episode_steps=arg.max_t)
eval_py_env = suite_gym.load(env_name,
discount=arg.gamma,
max_episode_steps=arg.max_t)
train_env = tf_py_environment.TFPyEnvironment(train_py_env)
eval_env = tf_py_environment.TFPyEnvironment(eval_py_env)
print("env loaded")
train_dir = os.path.join(arg.root_dir, 'network_weights')
eval_dir = os.path.join(arg.root_dir, 'eval')
train_env.reset()
fc_layer_params = (arg.hidden_size_1, )
optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=arg.lr)
train_step_counter = tf.compat.v2.Variable(0)
categorical_q_net = CategoricalQNetwork(train_env.observation_spec(),
train_env.action_spec(),
fc_layer_params=fc_layer_params)
agent = categorical_dqn_agent.CategoricalDqnAgent(
train_env.time_step_spec(),
train_env.action_spec(),
categorical_q_network=categorical_q_net,
optimizer=optimizer,
epsilon_greedy=arg.eps_start)
train_metrics = [
tf_metrics.NumberOfEpisodes(),
tf_metrics.EnvironmentSteps(),
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
global_step = tf.compat.v1.train.get_or_create_global_step()
train_checkpointer = common.Checkpointer(ckpt_dir=train_dir,
agent=tf_agent,
global_step=global_step,
metrics=metric_utils.MetricsGroup(
train_metrics,
'train_metrics'))
if arg.continue_training == False:
tf_agent.initialize()
if os.path.exists("network_weights/*"):
os.remove("network_weights/*")
else:
print("Continue Training")
train_checkpointer.initialize_or_restore()
print("ready to go")
eval_policy = tf_agent.policy
collect_policy = tf_agent.collect_policy
random_policy = random_tf_policy.RandomTFPolicy(train_env.time_step_spec(),
train_env.action_spec())
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
data_spec=tf_agent.collect_data_spec,
batch_size=train_env.batch_size,
max_length=arg.buffer_size)
tf_agent.collect_data_spec
tf_agent.collect_data_spec._fields
collect_data(train_env,
random_policy,
replay_buffer,
steps=arg.learn_start,
max_t=40)
print("create dataset")
dataset = replay_buffer.as_dataset(num_parallel_calls=3,
sample_batch_size=arg.batch_size,
num_steps=2).prefetch(3)
iterator = iter(dataset)
# (Optional) Optimize by wrapping some of the code in a graph using TF function.
tf_agent.train = common.function(tf_agent.train)
# Reset the train step
tf_agent.train_step_counter.assign(0)
avg_return = compute_avg_return(eval_env, tf_agent.policy,
arg.num_eval_episodes)
returns = [avg_return]
returns_average = [avg_return]
train_loss_average = [1]
score = 0
scores_window = deque(maxlen=100) # last 100 scores
total_train_loss = deque(maxlen=100) # last 100 scores
train(arg, tf_agent, train_env, eval_env, replay_buffer, iterator,
train_checkpointer)
categorical_q_net = categorical_q_network.CategoricalQNetwork(
train_env.observation_spec(),
train_env.action_spec(),
num_atoms=num_atoms,
fc_layer_params=fc_layer_params)
optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=learning_rate)
train_step_counter = tf.compat.v2.Variable(0)
agent = categorical_dqn_agent.CategoricalDqnAgent(
train_env.time_step_spec(),
train_env.action_spec(),
categorical_q_network=categorical_q_net,
optimizer=optimizer,
min_q_value=min_q_value,
max_q_value=max_q_value,
n_step_update=n_step_update,
td_errors_loss_fn=common.element_wise_squared_loss,
gamma=gamma,
train_step_counter=train_step_counter)
agent.initialize()
# setup the policies
eval_policy = agent.policy # The main policy that is used for evaluation and deployment
collect_policy = agent.collect_policy # A second policy that is used for data collection
random_policy = random_tf_policy.RandomTFPolicy(train_env.time_step_spec(),
train_env.action_spec())
# Data Collection