class DeepQAgent(object):
"""
Implementation of Deep Q Neural Network agent like in:
Nature 518. "Human-level control through deep reinforcement learning" (Mnih & al. 2015)
"""
def __init__(self, input_shape, nb_actions,
gamma=0.99, explorer=LinearEpsilonAnnealingExplorer(1, 0.1, 1000000),
learning_rate=0.00025, momentum=0.95, minibatch_size=32,
memory_size=500000, train_after=10000, train_interval=4, target_update_interval=10000,
monitor=True):
self.input_shape = input_shape
self.nb_actions = nb_actions
self.gamma = gamma
self._train_after = train_after
self._train_interval = train_interval
self._target_update_interval = target_update_interval
self._explorer = explorer
self._minibatch_size = minibatch_size
self._history = History(input_shape)
self._memory = ReplayMemory(memory_size, input_shape[1:], 4)
self._num_actions_taken = 0
# Metrics accumulator
self._episode_rewards, self._episode_q_means, self._episode_q_stddev = [], [], []
# Action Value model (used by agent to interact with the environment)
with default_options(activation=relu, init=he_uniform()):
self._action_value_net = Sequential([
Convolution2D((8, 8), 16, strides=4),
Convolution2D((4, 4), 32, strides=2),
Convolution2D((3, 3), 32, strides=1),
Dense(256, init=he_uniform(scale=0.01)),
Dense(nb_actions, activation=None, init=he_uniform(scale=0.01))
])
self._action_value_net.update_signature(Tensor[input_shape])
# Target model used to compute the target Q-values in training, updated
# less frequently for increased stability.
self._target_net = self._action_value_net.clone(CloneMethod.freeze)
# Function computing Q-values targets as part of the computation graph
@Function
@Signature(post_states=Tensor[input_shape], rewards=Tensor[()], terminals=Tensor[()])
def compute_q_targets(post_states, rewards, terminals):
return element_select(
terminals,
rewards,
gamma * reduce_max(self._target_net(post_states), axis=0) + rewards,
)
# Define the loss, using Huber Loss (more robust to outliers)
@Function
@Signature(pre_states=Tensor[input_shape], actions=Tensor[nb_actions],
post_states=Tensor[input_shape], rewards=Tensor[()], terminals=Tensor[()])
def criterion(pre_states, actions, post_states, rewards, terminals):
# Compute the q_targets
q_targets = compute_q_targets(post_states, rewards, terminals)
# actions is a 1-hot encoding of the action done by the agent
q_acted = reduce_sum(self._action_value_net(pre_states) * actions, axis=0)
# Define training criterion as the Huber Loss function
return huber_loss(q_targets, q_acted, 1.0)
# Adam based SGD
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
m_schedule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
l_sgd = adam(self._action_value_net.parameters, lr_schedule,
momentum=m_schedule, variance_momentum=vm_schedule)
self._metrics_writer = TensorBoardProgressWriter(freq=1, log_dir='metrics', model=criterion) if monitor else None
self._learner = l_sgd
self._trainer = Trainer(criterion, (criterion, None), l_sgd, self._metrics_writer)
def act(self, state):
""" This allows the agent to select the next action to perform in regard of the current state of the environment.
It follows the terminology used in the Nature paper.
Attributes:
state (Tensor[input_shape]): The current environment state
Returns: Int >= 0 : Next action to do
"""
# Append the state to the short term memory (ie. History)
self._history.append(state)
# If policy requires agent to explore, sample random action
if self._explorer.is_exploring(self._num_actions_taken):
action = self._explorer(self.nb_actions)
else:
# Use the network to output the best action
env_with_history = self._history.value
q_values = self._action_value_net.eval(
# Append batch axis with only one sample to evaluate
env_with_history.reshape((1,) + env_with_history.shape)
)
self._episode_q_means.append(np.mean(q_values))
self._episode_q_stddev.append(np.std(q_values))
# Return the value maximizing the expected reward
action = q_values.argmax()
# Keep track of interval action counter
self._num_actions_taken += 1
return action
def observe(self, old_state, action, reward, done):
""" This allows the agent to observe the output of doing the action it selected through act() on the old_state
Attributes:
old_state (Tensor[input_shape]): Previous environment state
action (int): Action done by the agent
reward (float): Reward for doing this action in the old_state environment
done (bool): Indicate if the action has terminated the environment
"""
self._episode_rewards.append(reward)
# If done, reset short term memory (ie. History)
if done:
# Plot the metrics through Tensorboard and reset buffers
if self._metrics_writer is not None:
self._plot_metrics()
self._episode_rewards, self._episode_q_means, self._episode_q_stddev = [], [], []
# Reset the short term memory
self._history.reset()
# Append to long term memory
self._memory.append(old_state, action, reward, done)
def train(self):
""" This allows the agent to train itself to better understand the environment dynamics.
The agent will compute the expected reward for the state(t+1)
and update the expected reward at step t according to this.
The target expectation is computed through the Target Network, which is a more stable version
of the Action Value Network for increasing training stability.
The Target Network is a frozen copy of the Action Value Network updated as regular intervals.
"""
agent_step = self._num_actions_taken
if agent_step >= self._train_after:
if (agent_step % self._train_interval) == 0:
pre_states, actions, post_states, rewards, terminals = self._memory.minibatch(self._minibatch_size)
self._trainer.train_minibatch(
self._trainer.loss_function.argument_map(
pre_states=pre_states,
actions=Value.one_hot(actions.reshape(-1, 1).tolist(), self.nb_actions),
post_states=post_states,
rewards=rewards,
terminals=terminals
)
)
# Update the Target Network if needed
if (agent_step % self._target_update_interval) == 0:
self._target_net = self._action_value_net.clone(CloneMethod.freeze)
filename = "models\model%d" % agent_step
self._trainer.save_checkpoint(filename)
def _plot_metrics(self):
"""Plot current buffers accumulated values to visualize agent learning
"""
if len(self._episode_q_means) > 0:
mean_q = np.asscalar(np.mean(self._episode_q_means))
self._metrics_writer.write_value('Mean Q per ep.', mean_q, self._num_actions_taken)
if len(self._episode_q_stddev) > 0:
std_q = np.asscalar(np.mean(self._episode_q_stddev))
self._metrics_writer.write_value('Mean Std Q per ep.', std_q, self._num_actions_taken)
self._metrics_writer.write_value('Sum rewards per ep.', sum(self._episode_rewards), self._num_actions_taken)
class DeepQAgent(object):
"""
Implementation of Deep Q Neural Network agent like in:
Nature 518. "Human-level control through deep reinforcement learning" (Mnih & al. 2015)
"""
def __init__(self, input_shape, nb_actions,
gamma=0.95, explorer=LinearEpsilonAnnealingExplorer(1, 0.1, 100000),
learning_rate=0.01, momentum=0.8, minibatch_size=16,
memory_size=15000, train_after=100, train_interval=100, target_update_interval=500,
monitor=True):
self.input_shape = input_shape
self.nb_actions = nb_actions
self.gamma = gamma
self._train_after = train_after
self._train_interval = train_interval
self._target_update_interval = target_update_interval
self._explorer = explorer
self._minibatch_size = minibatch_size
self._history = History(input_shape)
self._memory = ReplayMemory(memory_size, input_shape[1:], 4)
self._num_actions_taken = 0
self._num_trains = 0
# Metrics accumulator
self._episode_rewards, self._episode_q_means, self._episode_q_stddev = [], [], []
'''
# Action Value model (used by agent to interact with the environment)
with default_options(activation=relu, init=he_uniform()):
self._action_value_net = Sequential([
Convolution2D((8, 8), 16, strides=4),
Convolution2D((4, 4), 32, strides=2),
Convolution2D((3, 3), 32, strides=1),
Dense(256, init=he_uniform(scale=0.01)),
Dense(nb_actions, activation=None, init=he_uniform(scale=0.01))
])
'''
with default_options(activation=relu, init=he_uniform()):
self._action_value_net = Sequential([
Dense(7, init=he_uniform(scale=0.01)),
Dense(8, init=he_uniform(scale=0.01)),
#Dense(16, init=he_uniform(scale=0.01)),
#Dense(32, init=he_uniform(scale=0.01)),
Dense(nb_actions, activation=None, init=he_uniform(scale=0.01))
])
self._action_value_net.update_signature(Tensor[input_shape])
# Target model used to compute the target Q-values in training, updated
# less frequently for increased stability.
self._target_net = self._action_value_net.clone(CloneMethod.freeze)
# Function computing Q-values targets as part of the computation graph
@Function
@Signature(post_states=Tensor[input_shape], rewards=Tensor[()], terminals=Tensor[()])
def compute_q_targets(post_states, rewards, terminals):
return element_select(
terminals,
rewards,
gamma * reduce_max(self._target_net(post_states), axis=0) + rewards,
)
# Define the loss, using Huber Loss (more robust to outliers)
@Function
@Signature(pre_states=Tensor[input_shape], actions=Tensor[nb_actions],
post_states=Tensor[input_shape], rewards=Tensor[()], terminals=Tensor[()])
def criterion(pre_states, actions, post_states, rewards, terminals):
# Compute the q_targets
q_targets = compute_q_targets(post_states, rewards, terminals)
# actions is a 1-hot encoding of the action done by the agent
q_acted = reduce_sum(self._action_value_net(pre_states) * actions, axis=0)
# Define training criterion as the Huber Loss function
return huber_loss(q_targets, q_acted, 1.0)
# Adam based SGD
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
m_schedule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
l_sgd = adam(self._action_value_net.parameters, lr_schedule,
momentum=m_schedule, variance_momentum=vm_schedule)
self._metrics_writer = TensorBoardProgressWriter(freq=1, log_dir='metrics', model=criterion) if monitor else None
self._learner = l_sgd
self._trainer = Trainer(criterion, (criterion, None), l_sgd, self._metrics_writer)
#self._trainer.restore_from_checkpoint('models/oldmodels/model800000')
def act(self, state):
""" This allows the agent to select the next action to perform in regard of the current state of the environment.
It follows the terminology used in the Nature paper.
Attributes:
state (Tensor[input_shape]): The current environment state
Returns: Int >= 0 : Next action to do
"""
# Append the state to the short term memory (ie. History)
self._history.append(state)
# If policy requires agent to explore, sample random action
if self._explorer.is_exploring(self._num_actions_taken):
action = self._explorer(self.nb_actions)
else:
# Use the network to output the best action
env_with_history = self._history.value
q_values = self._action_value_net.eval(
# Append batch axis with only one sample to evaluate
env_with_history.reshape((1,) + env_with_history.shape)
)
self._episode_q_means.append(np.mean(q_values))
self._episode_q_stddev.append(np.std(q_values))
# Return the value maximizing the expected reward
action = q_values.argmax()
# Keep track of interval action counter
self._num_actions_taken += 1
#print(self._num_actions_taken)
return action
def observe(self, old_state, action, reward, done):
""" This allows the agent to observe the output of doing the action it selected through act() on the old_state
Attributes:
old_state (Tensor[input_shape]): Previous environment state
action (int): Action done by the agent
reward (float): Reward for doing this action in the old_state environment
done (bool): Indicate if the action has terminated the environment
"""
self._episode_rewards.append(reward)
# If done, reset short term memory (ie. History)
if done:
# Plot the metrics through Tensorboard and reset buffers
if self._metrics_writer is not None:
self._plot_metrics()
self._episode_rewards, self._episode_q_means, self._episode_q_stddev = [], [], []
# Reset the short term memory
self._history.reset()
# Append to long term memory
self._memory.append(old_state, action, reward, done)
def train(self):
""" This allows the agent to train itself to better understand the environment dynamics.
The agent will compute the expected reward for the state(t+1)
and update the expected reward at step t according to this.
The target expectation is computed through the Target Network, which is a more stable version
of the Action Value Network for increasing training stability.
The Target Network is a frozen copy of the Action Value Network updated as regular intervals.
"""
agent_step = self._num_actions_taken
if agent_step >= self._train_after:
#if (agent_step % self._train_interval) == 0:
print('\nTraining minibatch\n')
client.setCarControls(zero_controls)
pre_states, actions, post_states, rewards, terminals = self._memory.minibatch(self._minibatch_size)
self._trainer.train_minibatch(
self._trainer.loss_function.argument_map(
pre_states=pre_states,
actions=Value.one_hot(actions.reshape(-1, 1).tolist(), self.nb_actions),
post_states=post_states,
rewards=rewards,
terminals=terminals
)
)
self._num_trains += 1
# Update the Target Network if needed
if self._num_trains % 20 == 0:
print('updating network')
self._target_net = self._action_value_net.clone(CloneMethod.freeze)
filename = dirname+"\model%d" % agent_step
self._trainer.save_checkpoint(filename)
def _plot_metrics(self):
"""Plot current buffers accumulated values to visualize agent learning
"""
if len(self._episode_q_means) > 0:
mean_q = np.asscalar(np.mean(self._episode_q_means))
self._metrics_writer.write_value('Mean Q per ep.', mean_q, self._num_actions_taken)
if len(self._episode_q_stddev) > 0:
std_q = np.asscalar(np.mean(self._episode_q_stddev))
self._metrics_writer.write_value('Mean Std Q per ep.', std_q, self._num_actions_taken)
self._metrics_writer.write_value('Sum rewards per ep.', sum(self._episode_rewards), self._num_actions_taken)
class DQAgent(object):
"""docstring for DQAgent""" ############should modify @!@!
def __init__(self, input_shape, nb_actions,
gamma=0.99, explorer=LinearEpsilonAnnealingExplorer(1, 0.1, 1000000),
learning_rate=0.00025, momentum=0.95, minibatch_size=32,
memory_size=500000, train_after=10000, train_interval=4,
target_update_interval=10000, monitor=True):
self.input_shape = input_shape
self.nb_actions = nb_actions
self.gamma = gamma
self._train_after = train_after
self._train_interval = train_interval
self._target_update_interval = target_update_interval
self._explorer = explorer
self._minibatch_size = minibatch_size
self._history = History(input_shape)
self._memory = RepMem(memory_size, input_shape[1:], 4)
self._num_actions_taken = 0
self._episode_rewards, self._episode_q_means, self._episode_q_stddev = [], [], []
with default_options(activation=relu, init=he_uniform()):
self._action_value_net = Sequential([
Dense(input_shape, init=he_uniform(scale=0.01)),
Dense(input_shape),
Dense(nb_actions, activation=None, init=he_uniform(scale=0.01))])
self._action_value_net.update_signature(Tensor[input_shape])
self._target_net = self._action_value_net.clone(CloneMethod.freeze)
@Function
@Signature(post_states=Tensor[input_shape], rewards=Tensor[()], terminals=Tensor[()])
def compute_q_targets(post_states, rewards, terminals):
return element_select(
terminals,
rewards,
gamma * reduce_max(self._target_net(post_states), axis=0) + rewards,
)
@Function
@Signature(pre_states=Tensor[input_shape], actions=Tensor[nb_actions],
post_states=Tensor[input_shape], rewards=Tensor[()], terminals=Tensor[()])
def criterion(pre_states, actions, post_states, rewards, terminals):
q_targets = compute_q_targets(post_states, rewards, terminals)
q_acted = reduce_sum(self._action_value_net(pre_states) * actions, axis=0)
return huber_loss(q_targets, q_acted, 1.0)
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
m_schedule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
l_sgd = adam(self._action_value_net.parameters, lr_schedule,
momentum=m_schedule, variance_momentum=vm_schedule)
self._metrics_writer = TensorBoardProgressWriter(freq=1, log_dir='metrics', model=criterion) if monitor else None
self._learner = l_sgd
self._trainer = Trainer(criterion, (criterion, None), l_sgd, self._metrics_writer)
def act(self, state):
self._history.append(state)
if self._explorer.is_exploring(self._num_actions_taken):
action = self._explorer(self.nb_actions)
else:
env_with_history = self._history.value
q_values = self._action_value_net.eval(
env_with_history.reshape((1,) + env_with_history.shape)
)
self._episode_q_means.append(np.mean(q_values))
self._episode_q_stddev.append(np.std(q_values))
action = q_values.argmax()
self._num_actions_taken += 1
return action
def observe(self, old_state, action, reward, done):
self._episode_rewards.append(reward)
if done:
if self._metrics_writer is not None:
self._plot_metrics()
self._episode_rewards, self._episode_q_means, self._episode_q_stddev = [], [], []
self._history.reset()
self._memory.append(old_state, action, reward, done)
def train(self):
agent_step = self._num_actions_taken
if agent_step >= self._train_after:
if (agent_step % self._train_interval) == 0:
pre_states, actions, post_states, rewards, terminals = self._memory.minibatch(self._minibatch_size)
self._trainer.train_minibatch(
self._trainer.loss_function.argument_map(
pre_states=pre_states,
actions=Value.one_hot(actions.reshape(-1,1).tolist(), self.nb_actions),
post_states=post_states,
rewards=rewards,
terminals=terminals
)
)
if (agent_step % self._target_update_interval) == 0:
self._target_net = self._action_value_net.clone(CloneMethod.freeze)
filename = "model\model%d" % agent_step # save ???? not good at using %d
self._trainer.save_checkpoint(filename)
def _plot_metrics(self):
if len(self._episode_q_means) > 0:
mean_q = np.asscalar(np.mean(self._episode_q_means))
self._metrics_writer.write_value('Mean Q per ep.', mean_q, self._num_actions_taken)
if len(self._episode_q_stddev) > 0:
std_q = np.asscalar(np.mean(self._episode_q_stddev))
self._metrics_writer.write_value('Mean Std Q per ep.', std_q, self._num_actions_taken)
self._metrics_writer.write_value('Sum rewards per ep', sum(self._episode_rewards), self._num_actions_taken)
class DeepQAgent(object):
"""
Implementation of Deep Q Neural Network agent like in:
Nature 518. "Human-level control through deep reinforcement learning" (Mnih & al. 2015)
"""
def __init__(self,
input_shape,
nb_actions,
gamma=0.99,
explorer=LinearEpsilonAnnealingExplorer(1, 0.1, 1000000),
learning_rate=0.00025,
momentum=0.95,
minibatch_size=32,
memory_size=500000,
train_after=200000,
train_interval=4,
target_update_interval=10000,
monitor=True):
self.input_shape = input_shape
self.nb_actions = nb_actions
self.gamma = gamma
self._train_after = train_after
self._train_interval = train_interval
self._target_update_interval = target_update_interval
self._explorer = explorer
self._minibatch_size = minibatch_size
self._history = History(input_shape)
self._memory = ReplayMemory(memory_size, input_shape[1:], 4)
self._num_actions_taken = 0
# Metrics accumulator
self._episode_rewards, self._episode_q_means, self._episode_q_stddev = [], [], []
# Action Value model (used by agent to interact with the environment)
with default_options(activation=relu, init=he_uniform()):
self._action_value_net = Sequential([
Convolution2D((8, 8), 16, strides=4),
Convolution2D((4, 4), 32, strides=2),
Convolution2D((3, 3), 32, strides=1),
Dense(256, init=he_uniform(scale=0.01)),
Dense(nb_actions, activation=None, init=he_uniform(scale=0.01))
])
self._action_value_net.update_signature(Tensor[input_shape])
# Target model used to compute the target Q-values in training, updated
# less frequently for increased stability.
self._target_net = self._action_value_net.clone(CloneMethod.freeze)
# Function computing Q-values targets as part of the computation graph
@Function
@Signature(post_states=Tensor[input_shape],
rewards=Tensor[()],
terminals=Tensor[()])
def compute_q_targets(post_states, rewards, terminals):
return element_select(
terminals,
rewards,
gamma * reduce_max(self._target_net(post_states), axis=0) +
rewards,
)
# Define the loss, using Huber Loss (more robust to outliers)
@Function
@Signature(pre_states=Tensor[input_shape],
actions=Tensor[nb_actions],
post_states=Tensor[input_shape],
rewards=Tensor[()],
terminals=Tensor[()])
def criterion(pre_states, actions, post_states, rewards, terminals):
# Compute the q_targets
q_targets = compute_q_targets(post_states, rewards, terminals)
# actions is a 1-hot encoding of the action done by the agent
q_acted = reduce_sum(self._action_value_net(pre_states) * actions,
axis=0)
# Define training criterion as the Huber Loss function
return huber_loss(q_targets, q_acted, 1.0)
# Adam based SGD
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
m_schedule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
l_sgd = adam(self._action_value_net.parameters,
lr_schedule,
momentum=m_schedule,
variance_momentum=vm_schedule)
self._metrics_writer = TensorBoardProgressWriter(
freq=1, log_dir='metrics', model=criterion) if monitor else None
self._learner = l_sgd
self._trainer = Trainer(criterion, (criterion, None), l_sgd,
self._metrics_writer)
def load(self, model_path):
self._trainer.restore_from_checkpoint(model_path)
def act(self, state, eval=False):
""" This allows the agent to select the next action to perform in regard of the current state of the environment.
It follows the terminology used in the Nature paper.
Attributes:
state (Tensor[input_shape]): The current environment state
Returns: Int >= 0 : Next action to do
"""
# Append the state to the short term memory (ie. History)
self._history.append(state)
# If policy requires agent to explore, sample random action
if self._explorer.is_exploring(self._num_actions_taken) and not eval:
action = self._explorer(self.nb_actions)
q_values = None
else:
# Use the network to output the best action
env_with_history = self._history.value
q_values = self._action_value_net.eval(
# Append batch axis with only one sample to evaluate
env_with_history.reshape((1, ) + env_with_history.shape))
self._episode_q_means.append(np.mean(q_values))
self._episode_q_stddev.append(np.std(q_values))
# Return the value maximizing the expected reward
action = q_values.argmax()
# Keep track of interval action counter
self._num_actions_taken += 1
return action, q_values
def observe(self, old_state, action, reward, done):
""" This allows the agent to observe the output of doing the action it selected through act() on the old_state
Attributes:
old_state (Tensor[input_shape]): Previous environment state
action (int): Action done by the agent
reward (float): Reward for doing this action in the old_state environment
done (bool): Indicate if the action has terminated the environment
"""
self._episode_rewards.append(reward)
# If done, reset short term memory (ie. History)
if done:
# Plot the metrics through Tensorboard and reset buffers
if self._metrics_writer is not None:
self._plot_metrics()
self._episode_rewards, self._episode_q_means, self._episode_q_stddev = [], [], []
# Reset the short term memory
self._history.reset()
# Append to long term memory
self._memory.append(old_state, action, reward, done)
def train(self, checkpoint_dir):
""" This allows the agent to train itself to better understand the environment dynamics.
The agent will compute the expected reward for the state(t+1)
and update the expected reward at step t according to this.
The target expectation is computed through the Target Network, which is a more stable version
of the Action Value Network for increasing training stability.
The Target Network is a frozen copy of the Action Value Network updated as regular intervals.
"""
agent_step = self._num_actions_taken
if agent_step >= self._train_after:
if (agent_step % self._train_interval) == 0:
#print('training... number of steps: {}'.format(agent_step))
pre_states, actions, post_states, rewards, terminals = self._memory.minibatch(
self._minibatch_size)
self._trainer.train_minibatch(
self._trainer.loss_function.argument_map(
pre_states=pre_states,
actions=Value.one_hot(
actions.reshape(-1, 1).tolist(), self.nb_actions),
post_states=post_states,
rewards=rewards,
terminals=terminals))
# Update the Target Network if needed
if (agent_step % self._target_update_interval) == 0:
self._target_net = self._action_value_net.clone(
CloneMethod.freeze)
filename = os.path.join(checkpoint_dir,
"models\model%d" % agent_step)
self._trainer.save_checkpoint(filename)
def _plot_metrics(self):
"""Plot current buffers accumulated values to visualize agent learning
"""
if len(self._episode_q_means) > 0:
mean_q = np.asscalar(np.mean(self._episode_q_means))
self._metrics_writer.write_value('Mean Q per ep.', mean_q,
self._num_actions_taken)
if len(self._episode_q_stddev) > 0:
std_q = np.asscalar(np.mean(self._episode_q_stddev))
self._metrics_writer.write_value('Mean Std Q per ep.', std_q,
self._num_actions_taken)
self._metrics_writer.write_value('Sum rewards per ep.',
sum(self._episode_rewards),
self._num_actions_taken)
def get_depth_image(self, client):
# get depth image from airsim
responses = client.simGetImages([
airsim.ImageRequest("RCCamera", airsim.ImageType.DepthPerspective,
True, False)
])
img1d = np.array(responses[0].image_data_float, dtype=np.float)
img1d = 255 / np.maximum(np.ones(img1d.size), img1d)
if img1d.size > 1:
img2d = np.reshape(img1d,
(responses[0].height, responses[0].width))
image = Image.fromarray(img2d)
im_final = np.array(image.resize((84, 84)).convert('L'))
im_final = im_final / 255.0
return im_final
return np.zeros((84, 84)).astype(float)
# Gets a coverage image from AirSim
def get_cov_image(self, coverage_map):
state, cov_reward = coverage_map.get_state_from_pose()
#state = self.coverage_map.get_map_scaled()
# debug only
#im = Image.fromarray(np.uint8(state))
#im.save("DistributedRL\\debug\\{}.png".format(time.time()))
# normalize state
state = state / 255.0
return state, cov_reward
t += 1
# Stack together all inputs, hidden states, action gradients, and rewards for this episode
epx = np.vstack(states)
epl = np.vstack(labels).astype(np.float32)
epr = np.vstack(rewards).astype(np.float32)
# Compute the discounted reward backwards through time.
discounted_epr = discount_rewards(epr)
# Train the critic to predict the discounted reward from the observation
critic_trainer.train_minibatch({
observations: epx,
critic_target: discounted_epr
})
baseline = critic.eval({observations: epx})
# Compute n-step targets
n_step_targets = compute_n_step_targets(epr, baseline[0])
# Compute the baselined returns: A = n_step_targets - b(s). Weight the gradients by this value.
baselined_returns = n_step_targets - baseline
# Keep a running estimate over the variance of of the discounted rewards
for r in baselined_returns:
running_variance.add(r[0, 0])
# Forward pass
arguments = {
observations: epx,
label: epl,
class LearningAgent(object):
def __init__(self,
state_dim,
action_dim,
gamma=0.99,
learning_rate=1e-4,
momentum=0.95):
self.state_dim = state_dim
self.action_dim = action_dim
self.gamma = gamma
with default_options(activation=relu, init=he_uniform()):
# Convolution filter counts were halved to save on memory, no gpu :(
self.model = Sequential([
Convolution2D((8, 8), 16, strides=4, name='conv1'),
Convolution2D((4, 4), 32, strides=2, name='conv2'),
Convolution2D((3, 3), 32, strides=1, name='conv3'),
Dense(256, init=he_uniform(scale=0.01), name='dense1'),
Dense(action_dim,
activation=None,
init=he_uniform(scale=0.01),
name='actions')
])
self.model.update_signature(Tensor[state_dim])
# Create the target model as a copy of the online model
self.target_model = None
self.update_target()
self.pre_states = input_variable(state_dim, name='pre_states')
self.actions = input_variable(action_dim, name='actions')
self.post_states = input_variable(state_dim, name='post_states')
self.rewards = input_variable((), name='rewards')
self.terminals = input_variable((), name='terminals')
self.is_weights = input_variable((), name='is_weights')
predicted_q = reduce_sum(self.model(self.pre_states) * self.actions,
axis=0)
# DQN - calculate target q values
# post_q = reduce_max(self.target_model(self.post_states), axis=0)
# DDQN - calculate target q values
online_selection = one_hot(
argmax(self.model(self.post_states), axis=0), self.action_dim)
post_q = reduce_sum(self.target_model(self.post_states) *
online_selection,
axis=0)
post_q = (1.0 - self.terminals) * post_q
target_q = stop_gradient(self.rewards + self.gamma * post_q)
# Huber loss
delta = 1.0
self.td_error = minus(predicted_q, target_q, name='td_error')
abs_error = abs(self.td_error)
errors = element_select(less(abs_error, delta),
square(self.td_error) * 0.5,
delta * (abs_error - 0.5 * delta))
loss = errors * self.is_weights
# Adam based SGD
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
m_scheule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
self._learner = adam(self.model.parameters,
lr_schedule,
m_scheule,
variance_momentum=vm_schedule)
self.writer = TensorBoardProgressWriter(log_dir='metrics',
model=self.model)
self.trainer = Trainer(self.model, (loss, None), [self._learner],
self.writer)
def act(self, state, epsilon):
"""
Selects an action to take based on the epsilon greedy method
:param state: The current state
:param epsilon: Determines the amount of exploration. (1 - full exploration, 0 - no exploration)
"""
if np.random.randn(1) < epsilon:
# Explore (random action)
return np.random.choice(self.action_dim)
else:
# Exploit (greedy action based on knowledge)
return self.model.eval(state).argmax()
def train(self, s, a, r, s_, t, w):
"""
Updates the network weights using the given minibatch data
:param s: Tensor[state_dim] Current state
:param a: Tensor[int] Action taken at state s
:param r: Tensor[float] State resulting from taking action a at state s
:param s_: Tensor[state_dim] Reward received for taking action a at state s
:param t: Tensor[boolean] True if s_ was a terminal state and false otherwise
:param w: Tensor[float] Importance sampling weights
"""
a = Value.one_hot(a.tolist(), self.action_dim)
td_error = self.trainer.train_minibatch(
{
self.pre_states: s,
self.actions: a,
self.rewards: r,
self.post_states: s_,
self.terminals: t,
self.is_weights: w
},
outputs=[self.td_error])
return td_error[0]
def update_target(self):
"""
Update the target network using the online network weights
"""
self.target_model = self.model.clone(CloneMethod.freeze)
def checkpoint(self, filename):
self.trainer.save_checkpoint(filename)
def save_model(self, filename):
self.model.save(filename)
class DeepQAgent(object):
"""
Implementation of Deep Q Neural Network agent like in:
Nature 518. "Human-level control through deep reinforcement learning" (Mnih & al. 2015)
"""
def __init__(self,
input_shape,
nb_actions,
gamma=0.99,
explorer=LinearEpsilonAnnealingExplorer(0.91, 0.1, 910000),
fixpolicy=LinearEpsilonAnnealingExplorer(0.5, 0.1, 100000),
learning_rate=0.00025,
momentum=0.95,
minibatch_size=32,
memory_size=500000,
train_after=10000,
train_interval=4,
target_update_interval=10000,
monitor=True):
self.input_shape = input_shape
self.nb_actions = nb_actions
self.gamma = gamma
self._train_after = train_after
self._train_interval = train_interval
self._target_update_interval = target_update_interval
self._explorer = explorer
self._fixpolicy = fixpolicy
self._minibatch_size = minibatch_size
self._history = History(input_shape)
print("input_shape:", input_shape)
print("input_shape[1:]", input_shape[1:])
self._memory = ReplayMemory(memory_size, input_shape[1:], 4)
self._num_actions_taken = 0
# Metrics accumulator
self._episode_rewards, self._episode_q_means, self._episode_q_stddev = [], [], []
# Action Value model (used by agent to interact with the environment)
with default_options(activation=relu, init=he_uniform()):
self._action_value_net = Sequential([
#Convolution2D((8, 8), 16, strides=4),
#Convolution2D((4, 4), 32, strides=2),
#Convolution2D((1, 1), 16, strides=1),
Dense(25, init=he_uniform(scale=0.01)),
Dense(nb_actions, activation=None, init=he_uniform(scale=0.01))
])
self._action_value_net.update_signature(Tensor[input_shape])
# Target model used to compute the target Q-values in training, updated
# less frequently for increased stability.
self._target_net = self._action_value_net.clone(CloneMethod.freeze)
# Function computing Q-values targets as part of the computation graph
@Function
@Signature(post_states=Tensor[input_shape],
rewards=Tensor[()],
terminals=Tensor[()])
def compute_q_targets(post_states, rewards, terminals):
return element_select(
terminals,
rewards,
gamma * reduce_max(self._target_net(post_states), axis=0) +
rewards,
)
# Define the loss, using Huber Loss (more robust to outliers)
@Function
@Signature(pre_states=Tensor[input_shape],
actions=Tensor[nb_actions],
post_states=Tensor[input_shape],
rewards=Tensor[()],
terminals=Tensor[()])
def criterion(pre_states, actions, post_states, rewards, terminals):
# Compute the q_targets
q_targets = compute_q_targets(post_states, rewards, terminals)
# actions is a 1-hot encoding of the action done by the agent
q_acted = reduce_sum(self._action_value_net(pre_states) * actions,
axis=0)
# Define training criterion as the Huber Loss function
return huber_loss(q_targets, q_acted, 1.0)
# Adam based SGD
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
m_schedule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
l_sgd = adam(self._action_value_net.parameters,
lr_schedule,
momentum=m_schedule,
variance_momentum=vm_schedule)
self._metrics_writer = TensorBoardProgressWriter(
freq=1, log_dir='metrics', model=criterion) if monitor else None
self._learner = l_sgd
self._trainer = Trainer(criterion, (criterion, None), l_sgd,
self._metrics_writer)
#self._trainer.restore_from_checkpoint("models_heuristic_no_image\model")
def act(self, state):
""" This allows the agent to select the next action to perform in regard of the current state of the environment.
It follows the terminology used in the Nature paper.
Attributes:
state (Tensor[input_shape]): The current environment state
Returns: Int >= 0 : Next action to do
"""
# Append the state to the short term memory (ie. History)
self._history.append(state)
#if True:
if self._fixpolicy.is_exploring(self._num_actions_taken):
diff_x = state[3] - state[0]
diff_y = state[4] - state[1]
diff_z = state[5] - state[2]
diff_arr = np.array([diff_x, diff_y, diff_z])
direction = np.argmax(np.absolute(diff_arr))
'''
abs_x = math.fabs(diff_x)
abs_y = math.fabs(diff_y)
abs_z = math.fabs(diff_z)
diff = [diff_x, diff_y, diff_z]
abs_diff = [abs_x, abs_y, abs_z]
print(diff, abs_diff)
m = max(abs_diff)
direction = diff.index(m)'''
print(diff_arr)
if diff_arr[direction] < 0:
fixaction = direction + 4
else:
fixaction = direction + 1
self._num_actions_taken += 1
return fixaction
# If policy requires agent to explore, sample random action
if self._explorer.is_exploring(self._num_actions_taken):
action = self._explorer(self.nb_actions)
else:
# Use the network to output the best action
env_with_history = self._history.value
q_values = self._action_value_net.eval(
# Append batch axis with only one sample to evaluate
env_with_history.reshape((1, ) + env_with_history.shape))
self._episode_q_means.append(np.mean(q_values))
self._episode_q_stddev.append(np.std(q_values))
# Return the value maximizing the expected reward
action = q_values.argmax()
# Keep track of interval action counter
self._num_actions_taken += 1
return action
def observe(self, old_state, action, reward, done):
""" This allows the agent to observe the output of doing the action it selected through act() on the old_state
Attributes:
old_state (Tensor[input_shape]): Previous environment state
action (int): Action done by the agent
reward (float): Reward for doing this action in the old_state environment
done (bool): Indicate if the action has terminated the environment
"""
self._episode_rewards.append(reward)
# If done, reset short term memory (ie. History)
if done:
# Plot the metrics through Tensorboard and reset buffers
if self._metrics_writer is not None:
self._plot_metrics()
self._episode_rewards, self._episode_q_means, self._episode_q_stddev = [], [], []
# Reset the short term memory
self._history.reset()
# Append to long term memory
self._memory.append(old_state, action, reward, done)
def train(self):
""" This allows the agent to train itself to better understand the environment dynamics.
The agent will compute the expected reward for the state(t+1)
and update the expected reward at step t according to this.
The target expectation is computed through the Target Network, which is a more stable version
of the Action Value Network for increasing training stability.
The Target Network is a frozen copy of the Action Value Network updated as regular intervals.
"""
agent_step = self._num_actions_taken
print("agent_step = ", agent_step)
#time.sleep(1)
if agent_step >= self._train_after:
if (agent_step % self._train_interval) == 0:
pre_states, actions, post_states, rewards, terminals = self._memory.minibatch(
self._minibatch_size)
self._trainer.train_minibatch(
self._trainer.loss_function.argument_map(
pre_states=pre_states,
actions=Value.one_hot(
actions.reshape(-1, 1).tolist(), self.nb_actions),
post_states=post_states,
rewards=rewards,
terminals=terminals))
# Update the Target Network if needed
if (agent_step % self._target_update_interval) == 0:
self._target_net = self._action_value_net.clone(
CloneMethod.freeze)
filename = "models_heuristic_no_image_less_exploration\model%d" % agent_step
print(
"$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$filename=",
filename)
self._trainer.save_checkpoint(filename)
#time.sleep(100)
def _plot_metrics(self):
global landing_count, episode_count
"""Plot current buffers accumulated values to visualize agent learning
"""
f = open('log__heuristic_no_image_less_exploration2', 'a+')
f.write('episode:' + str(episode_count) + ': exploration rate= ' +
str(self._explorer._rate) + ' heuristic fix rate= ' +
str(self._fixpolicy._rate) + '\n')
if len(self._episode_q_means) > 0:
mean_q = np.asscalar(np.mean(self._episode_q_means))
self._metrics_writer.write_value('Mean Q per ep.', mean_q,
self._num_actions_taken)
print('Mean Q per ep.', mean_q, self._num_actions_taken)
f.write('Mean Q per ep. ' + str(mean_q) + ' ' +
str(self._num_actions_taken) + '\n')
if len(self._episode_q_stddev) > 0:
std_q = np.asscalar(np.mean(self._episode_q_stddev))
self._metrics_writer.write_value('Mean Std Q per ep.', std_q,
self._num_actions_taken)
print('Mean Std Q per ep.', std_q, self._num_actions_taken)
f.write('Mean Std Q per ep. ' + str(std_q) + ' ' +
str(self._num_actions_taken) + '\n')
self._metrics_writer.write_value('Sum rewards per ep.',
sum(self._episode_rewards),
self._num_actions_taken)
print('Sum rewards per ep.', sum(self._episode_rewards),
self._num_actions_taken)
f.write('Sum rewards per ep. ' + str(sum(self._episode_rewards)) +
' ' + str(self._num_actions_taken) + '\n')
if landing_count > 0:
f.write('****************Success landing**********' +
str(landing_count) + '\n')
landing_count = 0
episode_count = 0
f.write('\n')