def try_to_progress(self):
# If in state (1) or (2) then return immediately
if self.status == Client.WAITING_FOR_EXAMPLE or self.status == Client.WAITING_FOR_ACTION:
return self.status
assert self.status == Client.WAITING_TO_RECEIVE
# If in state (3) then see if the message is available. If the message
# is available then return to waiting for an action or a new example.
if self.state is None:
feedback = self.server.receive_reset_feedback_nonblocking()
else:
feedback = self.server.receive_feedback_nonblocking()
if feedback is None:
return self.status
else:
if self.state is None:
# assert False, "state should not be none"
# Feedback is in response to reset
image, metadata = feedback
pose = int(metadata["y_angle"] / 15.0)
position_orientation = (metadata["x_pos"], metadata["z_pos"],
metadata["y_angle"])
self.state = AgentObservedState(
instruction=self.current_data_point.instruction,
config=self.config,
constants=self.constants,
start_image=image,
previous_action=None,
data_point=self.current_data_point)
# Waiting for action
self.status = Client.WAITING_FOR_ACTION
else:
# Feedback is in response to an action
image, reward, metadata = feedback
self.total_reward += reward
# Create a replay item unless it is forced
if not self.forced_stop:
all_rewards = self._get_all_rewards(metadata)
replay_item = ReplayMemoryItem(
self.state,
self.last_action,
reward,
log_prob=self.last_log_prob,
image_emb_seq=self.image_emb_seq,
factor_entropy=self.factor_entropy,
all_rewards=all_rewards)
self.batch_replay_items.append(replay_item)
# Update the agent state
self.state = self.state.update(
image,
self.last_action,
data_point=self.current_data_point)
if self.last_action == self.agent.action_space.get_stop_action_index(
):
# Update the scores based on meta_data
# self.meta_data_util.log_results(metadata)
if self.tensorboard is not None:
self.tensorboard.log_all_train_errors(
metadata["edit_dist_error"],
metadata["closest_dist_error"],
metadata["stop_dist_error"])
self.status = Client.WAITING_FOR_EXAMPLE
else:
if self.num_action >= self.max_num_actions:
# Send forced stop action and wait to receive
self._take_forced_stop()
self.status = Client.WAITING_TO_RECEIVE
else:
# Wait to take another action
self.status = Client.WAITING_FOR_ACTION
self.metadata = metadata
return self.status
def do_train_(shared_model,
config,
action_space,
meta_data_util,
constants,
train_dataset,
tune_dataset,
experiment,
experiment_name,
rank,
server,
logger,
model_type,
vocab,
use_pushover=False):
print("In training...")
launch_k_unity_builds([config["port"]],
"./simulators/house_3_elmer.x86_64")
server.initialize_server()
print("launched builds")
# Test policy
test_policy = gp.get_argmax_action
# torch.manual_seed(args.seed + rank)
if rank == 0: # client 0 creates a tensorboard server
tensorboard = Tensorboard(experiment_name)
else:
tensorboard = None
if use_pushover:
# pushover_logger = PushoverLogger(experiment_name)
pushover_logger = None
else:
pushover_logger = None
# Create a local model for rollouts
local_model = model_type(config, constants)
# local_model.train()
# Create the Agent
logger.log("STARTING AGENT")
tmp_agent = TmpHouseAgent(server=server,
model=local_model,
test_policy=test_policy,
action_space=action_space,
meta_data_util=meta_data_util,
config=config,
constants=constants)
logger.log("Created Agent...")
action_counts = [0] * action_space.num_actions()
max_epochs = constants["max_epochs"]
dataset_size = len(train_dataset)
tune_dataset_size = len(tune_dataset)
# Create the learner to compute the loss
learner = TmpSupervisedLearning(shared_model, local_model,
action_space, meta_data_util, config,
constants, tensorboard)
# TODO change 2 --- unity launch moved up
for epoch in range(1, max_epochs + 1):
for data_point_ix, data_point in enumerate(train_dataset):
# Sync with the shared model
# local_model.load_state_dict(shared_model.state_dict())
local_model.load_from_state_dict(shared_model.get_state_dict())
if (data_point_ix + 1) % 100 == 0:
logger.log("Done %d out of %d" %
(data_point_ix, dataset_size))
logger.log("Training data action counts %r" %
action_counts)
image, metadata = tmp_agent.server.reset_receive_feedback(
data_point)
# instruction = TmpSupervisedLearning.convert_text_to_indices(metadata["instruction"], vocab)
instruction = data_point.get_instruction()
# Pose and Orientation gone TODO change 3
state = AgentObservedState(instruction=instruction,
config=config,
constants=constants,
start_image=image,
previous_action=None,
data_point=data_point)
model_state = None
batch_replay_items = []
total_reward = 0
# trajectory = metadata["trajectory"]
trajectory = data_point.get_trajectory()[0:300]
for action in trajectory:
# Sample action using the policy
log_probabilities, model_state, image_emb_seq, state_feature = \
local_model.get_probs(state, model_state)
# Sample action from the probability
action_counts[action] += 1
# Send the action and get feedback
image, reward, metadata = tmp_agent.server.send_action_receive_feedback(
action)
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state,
action,
reward,
log_prob=log_probabilities)
batch_replay_items.append(replay_item)
# Update the agent state
# Pose and orientation gone, TODO change 4
state = state.update(image, action, data_point=data_point)
total_reward += reward
# Send final STOP action and get feedback
# Sample action using the policy
log_probabilities, model_state, image_emb_seq, state_feature = \
local_model.get_probs(state, model_state)
image, reward, metadata = tmp_agent.server.halt_and_receive_feedback(
)
total_reward += reward
# if tensorboard is not None:
# tensorboard.log_all_train_errors(
# metadata["edit_dist_error"], metadata["closest_dist_error"], metadata["stop_dist_error"])
# Store it in the replay memory list
replay_item = ReplayMemoryItem(
state,
action_space.get_stop_action_index(),
reward,
log_prob=log_probabilities)
batch_replay_items.append(replay_item)
# Update the scores based on meta_data
# self.meta_data_util.log_results(metadata)
# Perform update
if len(batch_replay_items) > 0: # 32
loss_val = learner.do_update(batch_replay_items)
# self.action_prediction_loss_calculator.predict_action(batch_replay_items)
# del batch_replay_items[:] # in place list clear
if tensorboard is not None:
# cross_entropy = float(learner.cross_entropy.data[0])
# tensorboard.log(cross_entropy, loss_val, 0)
num_actions = len(trajectory) + 1
tensorboard.log_scalar(
"loss_val", loss_val) # /float(num_actions))
entropy = float(
learner.entropy.data[0]) # /float(num_actions)
tensorboard.log_scalar("entropy", entropy)
ratio = float(learner.ratio.data[0])
tensorboard.log_scalar(
"Abs_objective_to_entropy_ratio", ratio)
if learner.action_prediction_loss is not None:
action_prediction_loss = float(
learner.action_prediction_loss.data[0])
learner.tensorboard.log_action_prediction_loss(
action_prediction_loss)
if learner.temporal_autoencoder_loss is not None:
temporal_autoencoder_loss = float(
learner.temporal_autoencoder_loss.data[0])
tensorboard.log_temporal_autoencoder_loss(
temporal_autoencoder_loss)
if learner.object_detection_loss is not None:
object_detection_loss = float(
learner.object_detection_loss.data[0])
tensorboard.log_object_detection_loss(
object_detection_loss)
if learner.symbolic_language_prediction_loss is not None:
symbolic_language_prediction_loss = float(
learner.symbolic_language_prediction_loss.
data[0])
tensorboard.log_scalar(
"sym_language_prediction_loss",
symbolic_language_prediction_loss)
if learner.goal_prediction_loss is not None:
goal_prediction_loss = float(
learner.goal_prediction_loss.data[0])
tensorboard.log_scalar("goal_prediction_loss",
goal_prediction_loss)
if learner.mean_factor_entropy is not None:
mean_factor_entropy = float(
learner.mean_factor_entropy.data[0])
tensorboard.log_factor_entropy_loss(
mean_factor_entropy)
# Save the model
local_model.save_model(experiment + "/contextual_bandit_" +
str(rank) + "_epoch_" + str(epoch))
logger.log("Training data action counts %r" % action_counts)
if tune_dataset_size > 0:
# Test on tuning data
print("Going for testing")
tmp_agent.test(tune_dataset,
vocab,
tensorboard=tensorboard,
logger=logger,
pushover_logger=pushover_logger)
print("Done testing")
def do_train(self, train_dataset, train_images, train_goal_location,
tune_dataset, tune_images, tune_goal_location,
experiment_name):
""" Perform training """
dataset_size = len(train_dataset)
tensorboard = self.tensorboard
for epoch in range(1, self.max_epoch + 1):
logging.info("Starting epoch %d", epoch)
# Test on tuning data
self.test(tune_dataset,
tune_images,
tune_goal_location,
tensorboard=tensorboard)
for data_point_ix, data_point in enumerate(train_dataset):
if (data_point_ix + 1) % 100 == 0:
logging.info("Done %d out of %d", data_point_ix,
dataset_size)
train_images_example = train_images[data_point_ix]
goal_location = train_goal_location[data_point_ix]
image = train_images_example[0]
model_state = None
state = AgentObservedState(
instruction=data_point.instruction,
config=self.config,
constants=self.constants,
start_image=image,
previous_action=None,
pose=None,
position_orientation=data_point.get_start_pos(),
data_point=data_point)
trajectory = data_point.get_trajectory()
traj_len = len(trajectory)
if self.only_first:
trajectory = trajectory[0:1]
batch_replay_items = []
for action_ix, action in enumerate(trajectory):
# Sample action using the policy
# Generate probabilities over actions
volatile = self.model.get_attention_prob(
state, model_state)
goal = goal_location[action_ix]
# Store it in the replay memory list
if not self.ignore_none or goal[0] is not None:
replay_item = ReplayMemoryItem(state,
action,
0,
volatile=volatile,
goal=goal)
batch_replay_items.append(replay_item)
if not self.only_first:
# Send the action and get feedback
image = train_images_example[action_ix + 1]
# Update the agent state
state = state.update(image,
action,
pose=None,
position_orientation=None,
data_point=data_point)
# Store it in the replay memory list
if not self.only_first:
goal = goal_location[traj_len]
if not self.ignore_none or goal[0] is not None:
volatile = self.model.get_attention_prob(
state, model_state)
replay_item = ReplayMemoryItem(
state,
self.action_space.get_stop_action_index(),
0,
volatile=volatile,
goal=goal)
batch_replay_items.append(replay_item)
# Perform update
if len(batch_replay_items) > 0:
loss_val = self.do_update(batch_replay_items)
if tensorboard is not None:
tensorboard.log_scalar("Loss", loss_val)
if self.goal_prediction_loss is not None:
goal_prediction_loss = float(
self.goal_prediction_loss.data[0])
tensorboard.log_scalar("goal_prediction_loss",
goal_prediction_loss)
if self.goal_prob is not None:
goal_prob = float(self.goal_prob.data[0])
tensorboard.log_scalar("goal_prob", goal_prob)
if self.object_detection_loss is not None:
object_detection_loss = float(
self.object_detection_loss.data[0])
tensorboard.log_scalar("object_detection_loss",
object_detection_loss)
if self.cross_entropy_loss is not None:
cross_entropy_loss = float(
self.cross_entropy_loss.data[0])
tensorboard.log_scalar("Cross_entropy_loss",
cross_entropy_loss)
if self.dist_loss is not None:
dist_loss = float(self.dist_loss.data[0])
tensorboard.log_scalar("Dist_loss", dist_loss)
# Save the model
self.model.save_model(experiment_name +
"/goal_prediction_supervised_epoch_" +
str(epoch))
def do_train_(simulator_file,
shared_model,
config,
action_space,
meta_data_util,
constants,
train_dataset,
tune_dataset,
experiment,
experiment_name,
rank,
server,
logger,
model_type,
use_pushover=False):
# Launch unity
launch_k_unity_builds([config["port"]], simulator_file)
server.initialize_server()
# Test policy
test_policy = gp.get_argmax_action
# torch.manual_seed(args.seed + rank)
if rank == 0: # client 0 creates a tensorboard server
tensorboard = Tensorboard(experiment_name)
else:
tensorboard = None
if use_pushover:
pushover_logger = PushoverLogger(experiment_name)
else:
pushover_logger = None
# Create a local model for rollouts
local_model = model_type(config, constants)
# local_model.train()
# Create the Agent
logger.log("STARTING AGENT")
agent = Agent(server=server,
model=local_model,
test_policy=test_policy,
action_space=action_space,
meta_data_util=meta_data_util,
config=config,
constants=constants)
logger.log("Created Agent...")
action_counts = [0] * action_space.num_actions()
max_epochs = constants["max_epochs"]
dataset_size = len(train_dataset)
tune_dataset_size = len(tune_dataset)
# Create the learner to compute the loss
learner = AsynchronousContextualBandit(shared_model, local_model,
action_space, meta_data_util,
config, constants, tensorboard)
for epoch in range(1, max_epochs + 1):
for data_point_ix, data_point in enumerate(train_dataset):
# Sync with the shared model
local_model.load_from_state_dict(shared_model.get_state_dict())
if (data_point_ix + 1) % 100 == 0:
logger.log("Done %d out of %d" %
(data_point_ix, dataset_size))
logger.log("Training data action counts %r" %
action_counts)
num_actions = 0
max_num_actions = constants["horizon"] + constants[
"max_extra_horizon"]
image, metadata = agent.server.reset_receive_feedback(
data_point)
state = AgentObservedState(instruction=data_point.instruction,
config=config,
constants=constants,
start_image=image,
previous_action=None,
data_point=data_point)
meta_data_util.start_state_update_metadata(state, metadata)
model_state = None
batch_replay_items = []
total_reward = 0
forced_stop = True
while num_actions < max_num_actions:
# Sample action using the policy
log_probabilities, model_state, image_emb_seq, volatile = \
local_model.get_probs(state, model_state)
probabilities = list(torch.exp(log_probabilities.data))[0]
# Sample action from the probability
action = gp.sample_action_from_prob(probabilities)
action_counts[action] += 1
if action == action_space.get_stop_action_index():
forced_stop = False
break
# Send the action and get feedback
image, reward, metadata = agent.server.send_action_receive_feedback(
action)
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state,
action,
reward,
log_prob=log_probabilities,
volatile=volatile)
batch_replay_items.append(replay_item)
# Update the agent state
state = state.update(image, action, data_point=data_point)
meta_data_util.state_update_metadata(state, metadata)
num_actions += 1
total_reward += reward
# Send final STOP action and get feedback
image, reward, metadata = agent.server.halt_and_receive_feedback(
)
total_reward += reward
if tensorboard is not None:
meta_data_util.state_update_metadata(tensorboard, metadata)
# Store it in the replay memory list
if not forced_stop:
replay_item = ReplayMemoryItem(
state,
action_space.get_stop_action_index(),
reward,
log_prob=log_probabilities,
volatile=volatile)
batch_replay_items.append(replay_item)
# Perform update
if len(batch_replay_items) > 0:
loss_val = learner.do_update(batch_replay_items)
if tensorboard is not None:
entropy = float(
learner.entropy.data[0]) / float(num_actions + 1)
tensorboard.log_scalar("loss", loss_val)
tensorboard.log_scalar("entropy", entropy)
tensorboard.log_scalar("total_reward", total_reward)
# Save the model
local_model.save_model(experiment + "/contextual_bandit_" +
str(rank) + "_epoch_" + str(epoch))
logger.log("Training data action counts %r" % action_counts)
if tune_dataset_size > 0:
# Test on tuning data
agent.test(tune_dataset,
tensorboard=tensorboard,
logger=logger,
pushover_logger=pushover_logger)
def do_train(self, train_dataset, train_images, tune_dataset, tune_images,
experiment_name):
""" Perform training """
dataset_size = len(train_dataset)
for epoch in range(1, self.max_epoch + 1):
logging.info("Starting epoch %d", epoch)
# Test on tuning data
self.calc_log_prob(tune_dataset,
tune_images,
tensorboard=self.tensorboard)
batch_replay_items = []
episodes_in_batch = 0
for data_point_ix, data_point in enumerate(train_dataset):
if (data_point_ix + 1) % 100 == 0:
logging.info("Done %d out of %d", data_point_ix,
dataset_size)
train_images_example = train_images[data_point_ix]
image = train_images_example[0]
symbolic_form = nav_drone_symbolic_instructions.get_nav_drone_symbolic_instruction_segment(
data_point)
model_state = None
state = AgentObservedState(instruction=data_point.instruction,
config=self.config,
constants=self.constants,
start_image=image,
previous_action=None,
pose=None,
position_orientation=None,
data_point=data_point)
trajectory = data_point.get_trajectory()
for action_ix, action in enumerate(trajectory):
# Sample action using the policy
# Generate probabilities over actions
log_probabilities, model_state, image_emb_seq = self.model.get_probs(
state, model_state)
# Send the action and get feedback
image = train_images_example[action_ix + 1]
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state,
action,
0,
log_prob=log_probabilities,
symbolic_text=symbolic_form,
image_emb_seq=image_emb_seq,
text_emb=model_state[0])
batch_replay_items.append(replay_item)
# Update the agent state
state = state.update(image,
action,
pose=None,
position_orientation=None,
data_point=data_point)
log_probabilities, model_state, image_emb_seq = self.model.get_probs(
state, model_state)
# Store it in the replay memory list
replay_item = ReplayMemoryItem(
state,
self.action_space.get_stop_action_index(),
0,
log_prob=log_probabilities,
symbolic_text=symbolic_form,
image_emb_seq=image_emb_seq,
text_emb=model_state[0])
batch_replay_items.append(replay_item)
# Perform update
episodes_in_batch += 1
if episodes_in_batch == 1:
episodes_in_batch = 0
loss_val = self.do_update(batch_replay_items)
del batch_replay_items[:] # in place list clear
self.tensorboard.log_scalar("loss", loss_val)
cross_entropy = float(self.cross_entropy.data[0])
self.tensorboard.log_scalar("cross_entropy", cross_entropy)
entropy = float(self.entropy.data[0])
self.tensorboard.log_scalar("entropy", entropy)
if self.action_prediction_loss is not None:
action_prediction_loss = float(
self.action_prediction_loss.data[0])
self.tensorboard.log_action_prediction_loss(
action_prediction_loss)
if self.temporal_autoencoder_loss is not None:
temporal_autoencoder_loss = float(
self.temporal_autoencoder_loss.data[0])
self.tensorboard.log_temporal_autoencoder_loss(
temporal_autoencoder_loss)
if self.object_detection_loss is not None:
object_detection_loss = float(
self.object_detection_loss.data[0])
self.tensorboard.log_object_detection_loss(
object_detection_loss)
if self.symbolic_language_prediction_loss is not None:
symbolic_language_prediction_loss = float(
self.symbolic_language_prediction_loss.data[0])
self.tensorboard.log_scalar(
"sym_language_prediction_loss",
symbolic_language_prediction_loss)
if self.mean_factor_entropy is not None:
mean_factor_entropy = float(
self.mean_factor_entropy.data[0])
self.tensorboard.log_factor_entropy_loss(
mean_factor_entropy)
# Save the model
self.model.save_model(experiment_name +
"/contextual_bandit_resnet_epoch_" +
str(epoch))
def do_train_(shared_model,
config,
action_space,
meta_data_util,
constants,
train_dataset,
tune_dataset,
experiment,
experiment_name,
rank,
server,
logger,
model_type,
use_pushover=False):
server.initialize_server()
# Test policy
test_policy = gp.get_argmax_action
# torch.manual_seed(args.seed + rank)
if rank == 0: # client 0 creates a tensorboard server
tensorboard = Tensorboard(experiment_name)
else:
tensorboard = None
if use_pushover:
pushover_logger = PushoverLogger(experiment_name)
else:
pushover_logger = None
# Create a local model for rollouts
local_model = model_type(config, constants)
# local_model.train()
# Create the Agent
logger.log("STARTING AGENT")
agent = Agent(server=server,
model=local_model,
test_policy=test_policy,
action_space=action_space,
meta_data_util=meta_data_util,
config=config,
constants=constants)
logger.log("Created Agent...")
action_counts = [0] * action_space.num_actions()
max_epochs = constants["max_epochs"]
dataset_size = len(train_dataset)
tune_dataset_size = len(tune_dataset)
# Create the learner to compute the loss
learner = AsynchronousAdvantageActorGAECritic(shared_model,
local_model,
action_space,
meta_data_util, config,
constants, tensorboard)
# Launch unity
launch_k_unity_builds([config["port"]],
"./simulators/NavDroneLinuxBuild.x86_64")
for epoch in range(1, max_epochs + 1):
learner.epoch = epoch
task_completion_accuracy = 0
mean_stop_dist_error = 0
stop_dist_errors = []
for data_point_ix, data_point in enumerate(train_dataset):
# Sync with the shared model
# local_model.load_state_dict(shared_model.state_dict())
local_model.load_from_state_dict(shared_model.get_state_dict())
if (data_point_ix + 1) % 100 == 0:
logger.log("Done %d out of %d" %
(data_point_ix, dataset_size))
logger.log("Training data action counts %r" %
action_counts)
num_actions = 0
max_num_actions = constants["horizon"] + constants[
"max_extra_horizon"]
image, metadata = agent.server.reset_receive_feedback(
data_point)
pose = int(metadata["y_angle"] / 15.0)
position_orientation = (metadata["x_pos"], metadata["z_pos"],
metadata["y_angle"])
state = AgentObservedState(
instruction=data_point.instruction,
config=config,
constants=constants,
start_image=image,
previous_action=None,
pose=pose,
position_orientation=position_orientation,
data_point=data_point)
state.goal = GoalPrediction.get_goal_location(
metadata, data_point, learner.image_height,
learner.image_width)
model_state = None
batch_replay_items = []
total_reward = 0
forced_stop = True
while num_actions < max_num_actions:
# Sample action using the policy
log_probabilities, model_state, image_emb_seq, volatile = \
local_model.get_probs(state, model_state)
probabilities = list(torch.exp(log_probabilities.data))[0]
# Sample action from the probability
action = gp.sample_action_from_prob(probabilities)
action_counts[action] += 1
# Generate goal
if config["do_goal_prediction"]:
goal = learner.goal_prediction_calculator.get_goal_location(
metadata, data_point, learner.image_height,
learner.image_width)
else:
goal = None
if action == action_space.get_stop_action_index():
forced_stop = False
break
# Send the action and get feedback
image, reward, metadata = agent.server.send_action_receive_feedback(
action)
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state,
action,
reward,
log_prob=log_probabilities,
volatile=volatile,
goal=goal)
batch_replay_items.append(replay_item)
# Update the agent state
pose = int(metadata["y_angle"] / 15.0)
position_orientation = (metadata["x_pos"],
metadata["z_pos"],
metadata["y_angle"])
state = state.update(
image,
action,
pose=pose,
position_orientation=position_orientation,
data_point=data_point)
state.goal = GoalPrediction.get_goal_location(
metadata, data_point, learner.image_height,
learner.image_width)
num_actions += 1
total_reward += reward
# Send final STOP action and get feedback
image, reward, metadata = agent.server.halt_and_receive_feedback(
)
total_reward += reward
if metadata["stop_dist_error"] < 5.0:
task_completion_accuracy += 1
mean_stop_dist_error += metadata["stop_dist_error"]
stop_dist_errors.append(metadata["stop_dist_error"])
if tensorboard is not None:
tensorboard.log_all_train_errors(
metadata["edit_dist_error"],
metadata["closest_dist_error"],
metadata["stop_dist_error"])
# Store it in the replay memory list
if not forced_stop:
replay_item = ReplayMemoryItem(
state,
action_space.get_stop_action_index(),
reward,
log_prob=log_probabilities,
volatile=volatile,
goal=goal)
batch_replay_items.append(replay_item)
# Update the scores based on meta_data
# self.meta_data_util.log_results(metadata)
# Perform update
if len(batch_replay_items) > 0: # 32:
loss_val = learner.do_update(batch_replay_items)
# self.action_prediction_loss_calculator.predict_action(batch_replay_items)
# del batch_replay_items[:] # in place list clear
if tensorboard is not None:
cross_entropy = float(learner.cross_entropy.data[0])
tensorboard.log(cross_entropy, loss_val, 0)
entropy = float(
learner.entropy.data[0]) / float(num_actions + 1)
v_value_loss_per_step = float(
learner.value_loss.data[0]) / float(num_actions +
1)
tensorboard.log_scalar("entropy", entropy)
tensorboard.log_scalar("total_reward", total_reward)
tensorboard.log_scalar("v_value_loss_per_step",
v_value_loss_per_step)
ratio = float(learner.ratio.data[0])
tensorboard.log_scalar(
"Abs_objective_to_entropy_ratio", ratio)
if learner.action_prediction_loss is not None:
action_prediction_loss = float(
learner.action_prediction_loss.data[0])
learner.tensorboard.log_action_prediction_loss(
action_prediction_loss)
if learner.temporal_autoencoder_loss is not None:
temporal_autoencoder_loss = float(
learner.temporal_autoencoder_loss.data[0])
tensorboard.log_temporal_autoencoder_loss(
temporal_autoencoder_loss)
if learner.object_detection_loss is not None:
object_detection_loss = float(
learner.object_detection_loss.data[0])
tensorboard.log_object_detection_loss(
object_detection_loss)
if learner.symbolic_language_prediction_loss is not None:
symbolic_language_prediction_loss = float(
learner.symbolic_language_prediction_loss.
data[0])
tensorboard.log_scalar(
"sym_language_prediction_loss",
symbolic_language_prediction_loss)
if learner.goal_prediction_loss is not None:
goal_prediction_loss = float(
learner.goal_prediction_loss.data[0])
tensorboard.log_scalar("goal_prediction_loss",
goal_prediction_loss)
# Save the model
local_model.save_model(experiment + "/contextual_bandit_" +
str(rank) + "_epoch_" + str(epoch))
logger.log("Training data action counts %r" % action_counts)
mean_stop_dist_error = mean_stop_dist_error / float(
len(train_dataset))
task_completion_accuracy = (task_completion_accuracy *
100.0) / float(len(train_dataset))
logger.log("Training: Mean stop distance error %r" %
mean_stop_dist_error)
logger.log("Training: Task completion accuracy %r " %
task_completion_accuracy)
bins = range(0, 80, 3) # range of distance
histogram, _ = np.histogram(stop_dist_errors, bins)
logger.log("Histogram of train errors %r " % histogram)
if tune_dataset_size > 0:
# Test on tuning data
agent.test(tune_dataset,
tensorboard=tensorboard,
logger=logger,
pushover_logger=pushover_logger)
def do_train_(shared_model,
config,
action_space,
meta_data_util,
constants,
train_dataset,
tune_dataset,
experiment,
experiment_name,
rank,
server,
logger,
model_type,
use_pushover=False):
server.initialize_server()
# Test policy
test_policy = gp.get_argmax_action
# torch.manual_seed(args.seed + rank)
if rank == 0: # client 0 creates a tensorboard server
tensorboard = Tensorboard(experiment_name)
else:
tensorboard = None
if use_pushover:
pushover_logger = PushoverLogger(experiment_name)
else:
pushover_logger = None
# Create a local model for rollouts
local_model = model_type(config, constants)
# Create the Agent
logger.log("STARTING AGENT")
agent = Agent(server=server,
model=local_model,
test_policy=test_policy,
action_space=action_space,
meta_data_util=meta_data_util,
config=config,
constants=constants)
logger.log("Created Agent...")
action_counts = [0] * action_space.num_actions()
max_epochs = constants["max_epochs"]
dataset_size = len(train_dataset)
tune_dataset_size = len(tune_dataset)
# Create the learner to compute the loss
learner = AsynchronousSupervisedLearning(shared_model, local_model,
action_space, meta_data_util,
config, constants,
tensorboard)
# Launch unity
launch_k_unity_builds([config["port"]],
"./simulators/NavDroneLinuxBuild.x86_64")
for epoch in range(1, max_epochs + 1):
learner.epoch = epoch
for data_point_ix, data_point in enumerate(train_dataset):
# Sync with the shared model
# local_model.load_state_dict(shared_model.state_dict())
local_model.load_from_state_dict(shared_model.get_state_dict())
if (data_point_ix + 1) % 100 == 0:
logger.log("Done %d out of %d" %
(data_point_ix, dataset_size))
logger.log("Training data action counts %r" %
action_counts)
num_actions = 0
trajectory = data_point.get_trajectory()
image, metadata = agent.server.reset_receive_feedback(
data_point)
pose = int(metadata["y_angle"] / 15.0)
position_orientation = (metadata["x_pos"], metadata["z_pos"],
metadata["y_angle"])
state = AgentObservedState(
instruction=data_point.instruction,
config=config,
constants=constants,
start_image=image,
previous_action=None,
pose=pose,
position_orientation=position_orientation,
data_point=data_point)
model_state = None
batch_replay_items = []
total_reward = 0
for action in trajectory:
# Sample action using the policy
log_probabilities, model_state, image_emb_seq, volatile = \
local_model.get_probs(state, model_state)
action_counts[action] += 1
# Generate goal
if config["do_goal_prediction"]:
goal = learner.goal_prediction_calculator.get_goal_location(
metadata, data_point, 8, 8)
# learner.goal_prediction_calculator.save_attention_prob(image, volatile)
# time.sleep(5)
else:
goal = None
# Send the action and get feedback
image, reward, metadata = agent.server.send_action_receive_feedback(
action)
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state,
action,
reward,
log_prob=log_probabilities,
volatile=volatile,
goal=goal)
batch_replay_items.append(replay_item)
# Update the agent state
pose = int(metadata["y_angle"] / 15.0)
position_orientation = (metadata["x_pos"],
metadata["z_pos"],
metadata["y_angle"])
state = state.update(
image,
action,
pose=pose,
position_orientation=position_orientation,
data_point=data_point)
num_actions += 1
total_reward += reward
# Sample action using the policy
log_probabilities, model_state, image_emb_seq, volatile = \
local_model.get_probs(state, model_state)
# Generate goal
if config["do_goal_prediction"]:
goal = learner.goal_prediction_calculator.get_goal_location(
metadata, data_point, 8, 8)
# learner.goal_prediction_calculator.save_attention_prob(image, volatile)
# time.sleep(5)
else:
goal = None
# Send final STOP action and get feedback
image, reward, metadata = agent.server.halt_and_receive_feedback(
)
total_reward += reward
if tensorboard is not None:
tensorboard.log_all_train_errors(
metadata["edit_dist_error"],
metadata["closest_dist_error"],
metadata["stop_dist_error"])
# Store it in the replay memory list
replay_item = ReplayMemoryItem(
state,
action_space.get_stop_action_index(),
reward,
log_prob=log_probabilities,
volatile=volatile,
goal=goal)
batch_replay_items.append(replay_item)
###########################################3
AsynchronousSupervisedLearning.save_goal(
batch_replay_items, data_point_ix, trajectory)
###########################################3
# Update the scores based on meta_data
# self.meta_data_util.log_results(metadata)
# Perform update
if len(batch_replay_items) > 0: # 32:
loss_val = learner.do_update(batch_replay_items)
# self.action_prediction_loss_calculator.predict_action(batch_replay_items)
# del batch_replay_items[:] # in place list clear
if tensorboard is not None:
cross_entropy = float(learner.cross_entropy.data[0])
tensorboard.log(cross_entropy, loss_val, 0)
entropy = float(
learner.entropy.data[0]) / float(num_actions + 1)
tensorboard.log_scalar("entropy", entropy)
tensorboard.log_scalar("total_reward", total_reward)
ratio = float(learner.ratio.data[0])
tensorboard.log_scalar(
"Abs_objective_to_entropy_ratio", ratio)
if learner.action_prediction_loss is not None:
action_prediction_loss = float(
learner.action_prediction_loss.data[0])
learner.tensorboard.log_action_prediction_loss(
action_prediction_loss)
if learner.temporal_autoencoder_loss is not None:
temporal_autoencoder_loss = float(
learner.temporal_autoencoder_loss.data[0])
tensorboard.log_temporal_autoencoder_loss(
temporal_autoencoder_loss)
if learner.object_detection_loss is not None:
object_detection_loss = float(
learner.object_detection_loss.data[0])
tensorboard.log_object_detection_loss(
object_detection_loss)
if learner.symbolic_language_prediction_loss is not None:
symbolic_language_prediction_loss = float(
learner.symbolic_language_prediction_loss.
data[0])
tensorboard.log_scalar(
"sym_language_prediction_loss",
symbolic_language_prediction_loss)
if learner.goal_prediction_loss is not None:
goal_prediction_loss = float(
learner.goal_prediction_loss.data[0])
tensorboard.log_scalar("goal_prediction_loss",
goal_prediction_loss)
if learner.goal_prob is not None:
goal_prob = float(learner.goal_prob.data[0])
tensorboard.log_scalar("goal_prob", goal_prob)
if learner.mean_factor_entropy is not None:
mean_factor_entropy = float(
learner.mean_factor_entropy.data[0])
tensorboard.log_factor_entropy_loss(
mean_factor_entropy)
# Save the model
local_model.save_model(experiment + "/supervised_learning_" +
str(rank) + "_epoch_" + str(epoch))
logger.log("Training data action counts %r" % action_counts)
if tune_dataset_size > 0:
# Test on tuning data
agent.test_goal_prediction(tune_dataset,
tensorboard=tensorboard,
logger=logger,
pushover_logger=pushover_logger)
def do_train_forced_reading(self, agent, train_dataset, tune_dataset,
experiment_name):
""" Perform training """
assert isinstance(
agent, ReadPointerAgent
), "This learning algorithm works only with READPointerAgent"
dataset_size = len(train_dataset)
for epoch in range(1, self.max_epoch + 1):
logging.info("Starting epoch %d", epoch)
action_counts = dict()
action_counts[ReadPointerAgent.READ_MODE] = [0] * 2
action_counts[ReadPointerAgent.
ACT_MODE] = [0] * self.action_space.num_actions()
# Test on tuning data
agent.test_forced_reading(tune_dataset,
tensorboard=self.tensorboard)
batch_replay_items = []
total_reward = 0
episodes_in_batch = 0
for data_point_ix, data_point in enumerate(train_dataset):
if (data_point_ix + 1) % 100 == 0:
logging.info("Done %d out of %d", data_point_ix,
dataset_size)
logging.info("Training data action counts %r",
action_counts)
num_actions = 0
max_num_actions = len(data_point.get_trajectory())
max_num_actions += self.constants["max_extra_horizon"]
image, metadata = agent.server.reset_receive_feedback(
data_point)
oracle_segments = data_point.get_instruction_oracle_segmented()
pose = int(metadata["y_angle"] / 15.0)
state = AgentObservedState(instruction=data_point.instruction,
config=self.config,
constants=self.constants,
start_image=image,
previous_action=None,
pose=pose)
per_segment_budget = int(max_num_actions /
len(oracle_segments))
num_segment_actions = 0
mode = ReadPointerAgent.READ_MODE
current_segment_ix = 0
while True:
if mode == ReadPointerAgent.READ_MODE:
# Find the number of tokens to read for the gold segment
num_segment_size = len(
oracle_segments[current_segment_ix])
current_segment_ix += 1
for i in range(0, num_segment_size):
state = state.update_on_read()
mode = ReadPointerAgent.ACT_MODE
elif mode == ReadPointerAgent.ACT_MODE:
# Sample action using the policy
# Generate probabilities over actions
probabilities = list(
torch.exp(self.model.get_probs(state, mode).data))
# Use test policy to get the action
action = gp.sample_action_from_prob(probabilities)
action_counts[mode][action] += 1
# deal with act mode boundary conditions
if num_actions >= max_num_actions:
forced_stop = True
break
elif action == agent.action_space.get_stop_action_index(
) or num_segment_actions > per_segment_budget:
if state.are_tokens_left_to_be_read():
# reward = self._calc_reward_act_halt(state)
if metadata["error"] < 5.0:
reward = 1.0
else:
reward = -1.0
# Add to replay memory
replay_item = ReplayMemoryItem(
state,
agent.action_space.get_stop_action_index(),
reward, mode)
if action == agent.action_space.get_stop_action_index(
):
batch_replay_items.append(replay_item)
mode = ReadPointerAgent.READ_MODE
agent.server.force_goal_update()
state = state.update_on_act_halt()
num_segment_actions = 0
else:
if action == agent.action_space.get_stop_action_index(
):
forced_stop = False
else: # stopping due to per segment budget exhaustion
forced_stop = True
break
else:
image, reward, metadata = agent.server.send_action_receive_feedback(
action)
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state,
action,
reward,
mode=mode)
batch_replay_items.append(replay_item)
# Update the agent state
pose = int(metadata["y_angle"] / 15.0)
state = state.update(image, action, pose=pose)
num_actions += 1
num_segment_actions += 1
total_reward += reward
else:
raise AssertionError(
"Mode should be either read or act. Unhandled mode: "
+ str(mode))
assert mode == ReadPointerAgent.ACT_MODE, "Agent should end on Act Mode"
# Send final STOP action and get feedback
image, reward, metadata = agent.server.halt_and_receive_feedback(
)
total_reward += reward
# Store it in the replay memory list
if not forced_stop:
replay_item = ReplayMemoryItem(
state, agent.action_space.get_stop_action_index(),
reward, mode)
batch_replay_items.append(replay_item)
# Update the scores based on meta_data
# self.meta_data_util.log_results(metadata)
# Perform update
episodes_in_batch += 1
if episodes_in_batch == 1:
loss_val = self.do_update(batch_replay_items)
batch_replay_items = []
entropy_val = float(self.entropy.data[0])
self.tensorboard.log(entropy_val, loss_val, total_reward)
total_reward = 0
episodes_in_batch = 0
self.tensorboard.log_train_error(metadata["error"])
# Save the model
self.model.save_model(
experiment_name +
"/read_pointer_forced_reading_contextual_bandit_resnet_epoch_"
+ str(epoch))
logging.info("Training data action counts %r", action_counts)
def do_train_forced_reading(self, agent, train_dataset, tune_dataset,
experiment_name):
""" Perform training """
assert isinstance(
agent, ReadPointerAgent
), "This learning algorithm works only with READPointerAgent"
dataset_size = len(train_dataset)
for epoch in range(1, self.max_epoch + 1):
logging.info("Starting epoch %d", epoch)
total_cb_segments = 0
num_reached_acceptable_circle = 0
total_segments = 0
total_supervised_segments = 0
action_counts = dict()
action_counts[ReadPointerAgent.READ_MODE] = [0] * 2
action_counts[ReadPointerAgent.
ACT_MODE] = [0] * self.action_space.num_actions()
# Test on tuning data
agent.test_forced_reading(tune_dataset,
tensorboard=self.tensorboard)
batch_replay_items = []
total_reward = 0
episodes_in_batch = 0
for data_point_ix, data_point in enumerate(train_dataset):
if (data_point_ix + 1) % 100 == 0:
logging.info("Done %d out of %d", data_point_ix,
dataset_size)
logging.info(
"Contextual bandit segments %r, success %r per.",
total_cb_segments,
(num_reached_acceptable_circle * 100) /
float(max(1, total_cb_segments)))
logging.info("Num segments %r, Percent supervised %r",
total_segments,
(total_supervised_segments * 100) /
float(max(1, total_segments)))
logging.info("Training data action counts %r",
action_counts)
num_actions = 0
max_num_actions = len(data_point.get_trajectory())
max_num_actions += self.constants["max_extra_horizon"]
image, metadata = agent.server.reset_receive_feedback(
data_point)
oracle_segments = data_point.get_instruction_oracle_segmented()
pose = int(metadata["y_angle"] / 15.0)
state = AgentObservedState(instruction=data_point.instruction,
config=self.config,
constants=self.constants,
start_image=image,
previous_action=None,
pose=pose)
per_segment_budget = int(max_num_actions /
len(oracle_segments))
num_segment_actions = 0
trajectory_segments = data_point.get_sub_trajectory_list()
mode = ReadPointerAgent.READ_MODE
current_segment_ix = 0
num_supervised_rollout = self.rollin_policy.num_oracle_rollin_segments(
len(trajectory_segments))
total_segments += len(trajectory_segments)
while True:
if mode == ReadPointerAgent.READ_MODE:
# Find the number of tokens to read for the gold segment
num_segment_size = len(
oracle_segments[current_segment_ix])
current_segment_ix += 1
for i in range(0, num_segment_size):
state = state.update_on_read()
mode = ReadPointerAgent.ACT_MODE
total_segments += 1
elif mode == ReadPointerAgent.ACT_MODE:
if current_segment_ix <= num_supervised_rollout:
# Do supervised learning for this segment
for action in trajectory_segments[
current_segment_ix - 1]:
image, reward, metadata = agent.server.send_action_receive_feedback(
action)
# Store it in the replay memory list. Use reward of 1 as it is supervised learning
all_rewards = self._get_all_rewards(metadata)
replay_item = ReplayMemoryItem(
state,
action,
reward=1,
mode=mode,
all_rewards=all_rewards)
batch_replay_items.append(replay_item)
# Update the agent state
pose = int(metadata["y_angle"] / 15.0)
state = state.update(image, action, pose=pose)
num_actions += 1
total_reward += reward
# Change the segment
assert metadata[
"goal_dist"] < 5.0, "oracle segments out of acceptable circle"
if state.are_tokens_left_to_be_read():
mode = ReadPointerAgent.READ_MODE
# Jump to the next goal
agent.server.force_goal_update()
state = state.update_on_act_halt()
num_segment_actions = 0
else:
forced_stop = True
break
else:
# Do contextual bandit for this segment and future
# Generate probabilities over actions
probabilities = list(
torch.exp(
self.model.get_probs(state, mode).data))
# Sample an action from the distribution
action = gp.sample_action_from_prob(probabilities)
action_counts[mode][action] += 1
# deal with act mode boundary conditions
if num_actions >= max_num_actions:
break
elif action == agent.action_space.get_stop_action_index(
) or num_segment_actions > per_segment_budget:
within_acceptable_circle = metadata[
"goal_dist"] < 5.0
if within_acceptable_circle:
num_reached_acceptable_circle += 1
total_cb_segments += 1
if state.are_tokens_left_to_be_read():
if within_acceptable_circle:
if metadata["error"] < 5.0:
reward = 1.0
else:
reward = -1.0
# Add to replay memory
all_rewards = metadata["all_reward"]
replay_item = ReplayMemoryItem(
state,
agent.action_space.
get_stop_action_index(),
reward,
mode,
all_rewards=all_rewards)
batch_replay_items.append(replay_item)
mode = ReadPointerAgent.READ_MODE
# Jump to the next goal
agent.server.force_goal_update()
state = state.update_on_act_halt()
num_segment_actions = 0
else:
# No point going any further so break
break
else:
break
else:
image, reward, metadata = agent.server.send_action_receive_feedback(
action)
# Store it in the replay memory list
all_rewards = self._get_all_rewards(metadata)
replay_item = ReplayMemoryItem(
state,
action,
reward,
mode=mode,
all_rewards=all_rewards)
batch_replay_items.append(replay_item)
# Update the agent state
pose = int(metadata["y_angle"] / 15.0)
state = state.update(image, action, pose=pose)
num_actions += 1
num_segment_actions += 1
total_reward += reward
else:
raise AssertionError(
"Mode should be either read or act. Unhandled mode: "
+ str(mode))
assert mode == ReadPointerAgent.ACT_MODE, "Agent should end on Act Mode"
# Send final STOP action and get feedback
image, reward, metadata = agent.server.halt_and_receive_feedback(
)
total_reward += reward
# Update the scores based on meta_data
# self.meta_data_util.log_results(metadata)
# Perform update
episodes_in_batch += 1
if episodes_in_batch == 1:
loss_val = self.do_update(batch_replay_items)
batch_replay_items = []
entropy_val = float(self.entropy.data[0])
self.tensorboard.log(entropy_val, loss_val, total_reward)
total_reward = 0
episodes_in_batch = 0
if self.tensorboard is not None:
self.tensorboard.log_all_train_errors(
metadata["edit_dist_error"],
metadata["closest_dist_error"],
metadata["stop_dist_error"])
# Save the model
self.model.save_model(
experiment_name +
"/read_pointer_forced_reading_curriculum_contextual_bandit_epoch_"
+ str(epoch))
logging.info("Training data action counts %r", action_counts)
def do_train(self, agent, experiment_name):
""" Perform training """
print("in training")
for epoch in range(1, self.max_epoch + 1):
logging.info("Starting epoch %r", epoch)
# Test on tuning data
# switch instruction set to test
agent.server.env.switch_instructions_set('test')
agent.test(30, tensorboard=self.tensorboard)
agent.server.env.switch_instructions_set('train')
for i in range(0, 500):
batch_replay_items = []
num_actions = 0
total_reward = 0
instruction, image, metadata = agent.server.reset_receive_feedback(
)
state = AgentObservedState(instruction=instruction,
config=self.config,
constants=self.constants,
start_image=image,
previous_action=None)
model_state = None
while True:
# Sample action using the policy
# Generate probabilities over actions
log_probabilities, model_state, _, _ = self.model.get_probs(
state, model_state)
probabilities = list(torch.exp(log_probabilities.data))
# Use test policy to get the action
action = gp.sample_action_from_prob(probabilities[0])
# logging.info('Train: probabilities:' + str(probabilities[0].cpu().numpy()) + ' , action taken: ' + str(action))
# Send the action and get feedback
image, reward, done, metadata = agent.server.send_action_receive_feedback(
action, num_actions)
total_reward += reward
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state,
action,
reward,
log_prob=log_probabilities)
batch_replay_items.append(replay_item)
# Update the agent state
state = state.update(image, action)
num_actions += 1
if done:
break
# Perform update
loss_val = self.do_update(batch_replay_items)
entropy_val = float(self.entropy.data[0])
self.tensorboard.log(entropy_val, loss_val, total_reward)
# Save the model
self.model.save_model(experiment_name +
"/contextual_bandit_epoch_" + str(epoch))
def do_train_(shared_model,
config,
action_space,
meta_data_util,
args,
constants,
train_dataset,
tune_dataset,
experiment,
experiment_name,
rank,
server,
logger,
model_type,
use_pushover=False):
server.initialize_server()
# Test policy
test_policy = gp.get_argmax_action
# torch.manual_seed(args.seed + rank)
if rank == 0: # client 0 creates a tensorboard server
tensorboard = Tensorboard(experiment_name)
else:
tensorboard = None
if use_pushover:
pushover_logger = PushoverLogger(experiment_name)
else:
pushover_logger = None
# Create a local model for rollouts
local_model = model_type(args, config=config)
if torch.cuda.is_available():
local_model.cuda()
local_model.train()
# Create the Agent
logger.log("STARTING AGENT")
agent = Agent(server=server,
model=local_model,
test_policy=test_policy,
action_space=action_space,
meta_data_util=meta_data_util,
config=config,
constants=constants)
logger.log("Created Agent...")
action_counts = [0] * action_space.num_actions()
max_epochs = constants["max_epochs"]
dataset_size = len(train_dataset)
tune_dataset_size = len(tune_dataset)
# Create the learner to compute the loss
learner = AsynchronousContextualBandit(shared_model, local_model,
action_space, meta_data_util,
config, constants, tensorboard)
# Launch unity
launch_k_unity_builds([
config["port"]
], "/home/dipendra/Downloads/NavDroneLinuxBuild/NavDroneLinuxBuild.x86_64"
)
for epoch in range(1, max_epochs + 1):
if tune_dataset_size > 0:
# Test on tuning data
agent.test(tune_dataset,
tensorboard=tensorboard,
logger=logger,
pushover_logger=pushover_logger)
for data_point_ix, data_point in enumerate(train_dataset):
# Sync with the shared model
# local_model.load_state_dict(shared_model.state_dict())
local_model.load_from_state_dict(shared_model.get_state_dict())
if (data_point_ix + 1) % 100 == 0:
logging.info("Done %d out of %d", data_point_ix,
dataset_size)
logging.info("Training data action counts %r",
action_counts)
num_actions = 0
# max_num_actions = len(data_point.get_trajectory())
# max_num_actions += self.constants["max_extra_horizon"]
max_num_actions = constants["horizon"]
image, metadata = agent.server.reset_receive_feedback(
data_point)
pose = int(metadata["y_angle"] / 15.0)
position_orientation = (metadata["x_pos"], metadata["z_pos"],
metadata["y_angle"])
state = AgentObservedState(
instruction=data_point.instruction,
config=config,
constants=constants,
start_image=image,
previous_action=None,
pose=pose,
position_orientation=position_orientation,
data_point=data_point)
model_state = None
batch_replay_items = []
total_reward = 0
forced_stop = True
while num_actions < max_num_actions:
# Sample action using the policy
log_probabilities, model_state, image_emb_seq, state_feature = \
local_model.get_probs(state, model_state)
probabilities = list(torch.exp(log_probabilities.data))[0]
# Sample action from the probability
action = gp.sample_action_from_prob(probabilities)
action_counts[action] += 1
if action == action_space.get_stop_action_index():
forced_stop = False
break
# Send the action and get feedback
image, reward, metadata = agent.server.send_action_receive_feedback(
action)
# Store it in the replay memory list
rewards = learner.get_all_rewards(metadata)
replay_item = ReplayMemoryItem(state,
action,
reward,
log_prob=log_probabilities,
all_rewards=rewards)
batch_replay_items.append(replay_item)
# Update the agent state
pose = int(metadata["y_angle"] / 15.0)
position_orientation = (metadata["x_pos"],
metadata["z_pos"],
metadata["y_angle"])
state = state.update(
image,
action,
pose=pose,
position_orientation=position_orientation,
data_point=data_point)
num_actions += 1
total_reward += reward
# Send final STOP action and get feedback
image, reward, metadata = agent.server.halt_and_receive_feedback(
)
rewards = learner.get_all_rewards(metadata)
total_reward += reward
if tensorboard is not None:
tensorboard.log_all_train_errors(
metadata["edit_dist_error"],
metadata["closest_dist_error"],
metadata["stop_dist_error"])
# Store it in the replay memory list
if not forced_stop:
replay_item = ReplayMemoryItem(
state,
action_space.get_stop_action_index(),
reward,
log_prob=log_probabilities,
all_rewards=rewards)
batch_replay_items.append(replay_item)
# Update the scores based on meta_data
# self.meta_data_util.log_results(metadata)
# Perform update
if len(batch_replay_items) > 0:
loss_val = learner.do_update(batch_replay_items)
# self.action_prediction_loss_calculator.predict_action(batch_replay_items)
del batch_replay_items[:] # in place list clear
if tensorboard is not None:
cross_entropy = float(learner.cross_entropy.data[0])
tensorboard.log(cross_entropy, loss_val, 0)
entropy = float(learner.entropy.data[0])
tensorboard.log_scalar("entropy", entropy)
ratio = float(learner.ratio.data[0])
tensorboard.log_scalar(
"Abs_objective_to_entropy_ratio", ratio)
if learner.action_prediction_loss is not None:
action_prediction_loss = float(
learner.action_prediction_loss.data[0])
learner.tensorboard.log_action_prediction_loss(
action_prediction_loss)
if learner.temporal_autoencoder_loss is not None:
temporal_autoencoder_loss = float(
learner.temporal_autoencoder_loss.data[0])
tensorboard.log_temporal_autoencoder_loss(
temporal_autoencoder_loss)
if learner.object_detection_loss is not None:
object_detection_loss = float(
learner.object_detection_loss.data[0])
tensorboard.log_object_detection_loss(
object_detection_loss)
if learner.symbolic_language_prediction_loss is not None:
symbolic_language_prediction_loss = float(
learner.symbolic_language_prediction_loss.
data[0])
tensorboard.log_scalar(
"sym_language_prediction_loss",
symbolic_language_prediction_loss)
if learner.goal_prediction_loss is not None:
goal_prediction_loss = float(
learner.goal_prediction_loss.data[0])
tensorboard.log_scalar("goal_prediction_loss",
goal_prediction_loss)
if learner.mean_factor_entropy is not None:
mean_factor_entropy = float(
learner.mean_factor_entropy.data[0])
tensorboard.log_factor_entropy_loss(
mean_factor_entropy)
# Save the model
local_model.save_model(experiment + "/contextual_bandit_" +
str(rank) + "_epoch_" + str(epoch))
logging.info("Training data action counts %r", action_counts)
def do_train(self, agent, train_dataset, tune_dataset, experiment_name):
""" Perform training """
dataset_size = len(train_dataset)
for epoch in range(1, self.max_epoch + 1):
logging.info("Starting epoch %d", epoch)
action_counts = [0] * self.action_space.num_actions()
# Test on tuning data
self.test(agent, tune_dataset)
# self.print_confusion_matrix()
batch_replay_items = []
total_reward = 0
episodes_in_batch = 0
for data_point_ix, data_point in enumerate(train_dataset):
if (data_point_ix + 1) % 100 == 0:
logging.info("Done %d out of %d", data_point_ix, dataset_size)
logging.info("Training data action counts %r", action_counts)
num_actions = 0
image, metadata = agent.server.reset_receive_feedback(data_point)
pose = int(metadata["y_angle"] / 15.0)
position_orientation = (metadata["x_pos"], metadata["z_pos"],
metadata["y_angle"])
state = AgentObservedState(instruction=data_point.instruction,
config=self.config,
constants=self.constants,
start_image=image,
previous_action=None,
pose=pose,
position_orientation=position_orientation,
data_point=data_point)
trajectory = data_point.get_trajectory()
for action in trajectory:
action_counts[action] += 1
# Send the action and get feedback
image, reward, metadata = agent.server.send_action_receive_feedback(action)
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state, action, reward)
batch_replay_items.append(replay_item)
# Update the agent state
pose = int(metadata["y_angle"] / 15.0)
position_orientation = (metadata["x_pos"],
metadata["z_pos"],
metadata["y_angle"])
state = state.update(
image, action, pose=pose,
position_orientation=position_orientation,
data_point=data_point)
num_actions += 1
total_reward += reward
# Send final STOP action and get feedback
image, reward, metadata = agent.server.halt_and_receive_feedback()
total_reward += reward
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state, agent.action_space.get_stop_action_index(), reward)
batch_replay_items.append(replay_item)
# Perform update
episodes_in_batch += 1
if episodes_in_batch == 1:
loss_val = self.do_update(batch_replay_items)
batch_replay_items = []
# cross_entropy = float(self.cross_entropy.data[0])
# self.tensorboard.log(cross_entropy, loss_val, total_reward)
total_reward = 0
episodes_in_batch = 0
if self.tensorboard is not None:
self.tensorboard.log_all_train_errors(
metadata["edit_dist_error"], metadata["closest_dist_error"], metadata["stop_dist_error"])
if data_point_ix == 2000:
self.test(agent, tune_dataset)
if data_point_ix == 6000:
self.test(agent, tune_dataset)
# Save the model
self.model.save_model(experiment_name + "/object_detection_resnet_epoch_" + str(epoch))
logging.info("Training data action counts %r", action_counts)
def do_train(self, agent, train_dataset, tune_dataset, experiment_name):
""" Perform training """
for epoch in range(1, self.max_epoch + 1):
# Test on tuning data
agent.test(tune_dataset, tensorboard=self.tensorboard)
for data_point in train_dataset:
batch_replay_items = []
num_actions = 0
total_reward = 0
max_num_actions = len(data_point.get_trajectory())
max_num_actions += self.constants["max_extra_horizon"]
image, metadata = agent.server.reset_receive_feedback(data_point)
state = AgentObservedState(instruction=data_point.instruction,
config=self.config,
constants=self.constants,
start_image=image,
previous_action=None)
forced_stop = True
instruction = instruction_to_string(
data_point.get_instruction(), self.config)
print "TRAIN INSTRUCTION: %r" % instruction
print ""
while num_actions < max_num_actions:
# Sample action using the policy
# Generate probabilities over actions
probabilities = list(torch.exp(self.model.get_probs(state).data))
# Use test policy to get the action
action = gp.sample_action_from_prob(probabilities)
if action == agent.action_space.get_stop_action_index():
forced_stop = False
break
# Send the action and get feedback
image, reward, metadata = agent.server.send_action_receive_feedback(action)
total_reward += reward
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state, action, reward)
batch_replay_items.append(replay_item)
# Update the agent state
state = state.update(image, action)
num_actions += 1
# Send final STOP action and get feedback
image, reward, metadata = agent.server.halt_and_receive_feedback()
total_reward += reward
# Store it in the replay memory list
if not forced_stop:
replay_item = ReplayMemoryItem(state, agent.action_space.get_stop_action_index(), reward)
batch_replay_items.append(replay_item)
# Update the scores based on meta_data
# self.meta_data_util.log_results(metadata)
# Compute Q-values using sampled rollout
ReinforceLearning._set_q_val(batch_replay_items)
# Perform update
loss_val = self.do_update(batch_replay_items)
entropy_val = float(self.entropy.data[0])
self.tensorboard.log(entropy_val, loss_val, total_reward)
self.tensorboard.log_train_error(metadata["error"])
# Save the model
self.model.save_model(experiment_name + "/reinforce_epoch_" + str(epoch))
def do_train_(house_id, shared_model, config, action_space, meta_data_util, constants,
train_dataset, tune_dataset, experiment, experiment_name, rank, server,
logger, model_type, vocab, use_pushover=False):
logger.log("In Training...")
launch_k_unity_builds([config["port"]], "./house_" + str(house_id) + "_elmer.x86_64",
arg_str="--config ./AssetsHouse/config" + str(house_id) + ".json",
cwd="./simulators/house/")
logger.log("Launched Builds.")
server.initialize_server()
logger.log("Server Initialized.")
# Test policy
test_policy = gp.get_argmax_action
if rank == 0: # client 0 creates a tensorboard server
tensorboard = Tensorboard(experiment_name)
logger.log('Created Tensorboard Server.')
else:
tensorboard = None
if use_pushover:
pushover_logger = None
else:
pushover_logger = None
# Create a local model for rollouts
local_model = model_type(config, constants)
# local_model.train()
# Create the Agent
tmp_agent = TmpHouseAgent(server=server,
model=local_model,
test_policy=test_policy,
action_space=action_space,
meta_data_util=meta_data_util,
config=config,
constants=constants)
logger.log("Created Agent.")
action_counts = [0] * action_space.num_actions()
max_epochs = 100000 # constants["max_epochs"]
dataset_size = len(train_dataset)
tune_dataset_size = len(tune_dataset)
if tune_dataset_size > 0:
# Test on tuning data
tmp_agent.test(tune_dataset, vocab, tensorboard=tensorboard,
logger=logger, pushover_logger=pushover_logger)
# Create the learner to compute the loss
learner = TmpAsynchronousContextualBandit(shared_model, local_model, action_space, meta_data_util,
config, constants, tensorboard)
# TODO change 2 --- unity launch moved up
learner.logger = logger
for epoch in range(1, max_epochs + 1):
for data_point_ix, data_point in enumerate(train_dataset):
# Sync with the shared model
# local_model.load_state_dict(shared_model.state_dict())
local_model.load_from_state_dict(shared_model.get_state_dict())
if (data_point_ix + 1) % 100 == 0:
logger.log("Done %d out of %d" %(data_point_ix, dataset_size))
logger.log("Training data action counts %r" % action_counts)
num_actions = 0
max_num_actions = constants["horizon"]
max_num_actions += constants["max_extra_horizon"]
image, metadata = tmp_agent.server.reset_receive_feedback(data_point)
instruction = data_point.get_instruction()
# instruction_str = TmpAsynchronousContextualBandit.convert_indices_to_text(instruction, vocab)
# print("Instruction str is ", instruction_str)
# Pose and Orientation gone TODO change 3
state = AgentObservedState(instruction=instruction,
config=config,
constants=constants,
start_image=image,
previous_action=None,
data_point=data_point)
state.goal = learner.get_goal(metadata)
model_state = None
batch_replay_items = []
total_reward = 0
forced_stop = True
while num_actions < max_num_actions:
# logger.log("Training: Meta Data %r " % metadata)
# Sample action using the policy
log_probabilities, model_state, image_emb_seq, state_feature = \
local_model.get_probs(state, model_state)
probabilities = list(torch.exp(log_probabilities.data))[0]
# Sample action from the probability
action = gp.sample_action_from_prob(probabilities)
action_counts[action] += 1
if action == action_space.get_stop_action_index():
forced_stop = False
break
# Send the action and get feedback
image, reward, metadata = tmp_agent.server.send_action_receive_feedback(action)
# logger.log("Action is %r, Reward is %r Probability is %r " % (action, reward, probabilities))
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state, action, reward, log_prob=log_probabilities)
batch_replay_items.append(replay_item)
# Update the agent state
# Pose and orientation gone, TODO change 4
state = state.update(image, action, data_point=data_point)
state.goal = learner.get_goal(metadata)
num_actions += 1
total_reward += reward
# Send final STOP action and get feedback
image, reward, metadata = tmp_agent.server.halt_and_receive_feedback()
total_reward += reward
# Store it in the replay memory list
if not forced_stop:
# logger.log("Action is Stop, Reward is %r Probability is %r " % (reward, probabilities))
replay_item = ReplayMemoryItem(state, action_space.get_stop_action_index(),
reward, log_prob=log_probabilities)
batch_replay_items.append(replay_item)
# Update the scores based on meta_data
# self.meta_data_util.log_results(metadata)
# Perform update
if len(batch_replay_items) > 0: # 32
loss_val = learner.do_update(batch_replay_items)
if tensorboard is not None:
# cross_entropy = float(learner.cross_entropy.data[0])
# tensorboard.log(cross_entropy, loss_val, 0)
tensorboard.log_scalar("loss", loss_val)
entropy = float(learner.entropy.data[0])/float(num_actions + 1)
tensorboard.log_scalar("entropy", entropy)
ratio = float(learner.ratio.data[0])
tensorboard.log_scalar("Abs_objective_to_entropy_ratio", ratio)
tensorboard.log_scalar("total_reward", total_reward)
tensorboard.log_scalar("mean navigation error", metadata['mean-navigation-error'])
if learner.action_prediction_loss is not None:
action_prediction_loss = float(learner.action_prediction_loss.data[0])
learner.tensorboard.log_action_prediction_loss(action_prediction_loss)
if learner.temporal_autoencoder_loss is not None:
temporal_autoencoder_loss = float(learner.temporal_autoencoder_loss.data[0])
tensorboard.log_temporal_autoencoder_loss(temporal_autoencoder_loss)
if learner.object_detection_loss is not None:
object_detection_loss = float(learner.object_detection_loss.data[0])
tensorboard.log_object_detection_loss(object_detection_loss)
if learner.symbolic_language_prediction_loss is not None:
symbolic_language_prediction_loss = float(learner.symbolic_language_prediction_loss.data[0])
tensorboard.log_scalar("sym_language_prediction_loss", symbolic_language_prediction_loss)
if learner.goal_prediction_loss is not None:
goal_prediction_loss = float(learner.goal_prediction_loss.data[0])
tensorboard.log_scalar("goal_prediction_loss", goal_prediction_loss)
# Save the model
local_model.save_model(experiment + "/contextual_bandit_" + str(rank) + "_epoch_" + str(epoch))
logger.log("Training data action counts %r" % action_counts)
if tune_dataset_size > 0:
# Test on tuning data
tmp_agent.test(tune_dataset, vocab, tensorboard=tensorboard,
logger=logger, pushover_logger=pushover_logger)
def do_train(self,
train_dataset,
tune_dataset,
experiment_name,
save_best_model=False):
""" Perform training """
dataset_size = len(train_dataset)
tensorboard = self.tensorboard
# Test on tuning data with initialized model
mean_real_world_distance = self.test(tune_dataset,
tensorboard=tensorboard)
best_real_world_distance = mean_real_world_distance
for epoch in range(1, self.max_epoch + 1):
logging.info("Starting epoch %d", epoch)
batch_replay_items = []
best_real_world_distance = min(best_real_world_distance,
mean_real_world_distance)
for data_point_ix, data_point in enumerate(train_dataset):
if (data_point_ix + 1) % 100 == 0:
logging.info("Done %d out of %d", data_point_ix,
dataset_size)
model_state = None
state = AgentObservedState(instruction=data_point.instruction,
config=self.config,
constants=self.constants,
start_image=data_point.start_image,
previous_action=None,
pose=None,
position_orientation=None,
data_point=data_point)
# Generate attention probabilities
volatile = self.model.get_attention_prob(state, model_state)
row, col = data_point.goal_pixel
goal = row, col, row, col
# Store it in the replay memory list
if not self.ignore_none or goal[0] is not None:
replay_item = ReplayMemoryItem(state,
None,
0,
volatile=volatile,
goal=goal)
batch_replay_items.append(replay_item)
# Perform update
if len(batch_replay_items) > 0:
loss_val = self.do_update(batch_replay_items)
batch_replay_items = []
if tensorboard is not None:
tensorboard.log_scalar("Loss", loss_val)
if self.goal_prediction_loss is not None:
goal_prediction_loss = float(
self.goal_prediction_loss.data[0])
tensorboard.log_scalar("goal_prediction_loss",
goal_prediction_loss)
if self.goal_prob is not None:
goal_prob = float(self.goal_prob.data[0])
tensorboard.log_scalar("goal_prob", goal_prob)
if self.object_detection_loss is not None:
object_detection_loss = float(
self.object_detection_loss.data[0])
tensorboard.log_scalar("object_detection_loss",
object_detection_loss)
if self.cross_entropy_loss is not None:
cross_entropy_loss = float(
self.cross_entropy_loss.data[0])
tensorboard.log_scalar("Cross_entropy_loss",
cross_entropy_loss)
if self.dist_loss is not None:
dist_loss = float(self.dist_loss.data[0])
tensorboard.log_scalar("Dist_loss", dist_loss)
mean_real_world_distance = self.test(tune_dataset,
tensorboard=tensorboard)
# Save the model
if save_best_model:
if mean_real_world_distance < best_real_world_distance:
self.model.save_model(
experiment_name +
"/goal_prediction_single_supervised_epoch_" +
str(epoch))
else:
self.model.save_model(
experiment_name +
"/goal_prediction_single_supervised_epoch_" + str(epoch))
def do_train(self, agent, train_dataset, tune_dataset, experiment_name):
""" Perform training """
assert isinstance(
agent, ReadPointerAgent
), "This learning algorithm works only with READPointerAgent"
dataset_size = len(train_dataset)
for epoch in range(1, self.max_epoch + 1):
logging.info("Starting epoch %d", epoch)
action_counts = dict()
action_counts[ReadPointerAgent.READ_MODE] = [0] * 2
action_counts[ReadPointerAgent.
ACT_MODE] = [0] * self.action_space.num_actions()
# Test on tuning data
agent.test(tune_dataset, tensorboard=self.tensorboard)
batch_replay_items = []
total_reward = 0
episodes_in_batch = 0
for data_point_ix, data_point in enumerate(train_dataset):
if (data_point_ix + 1) % 100 == 0:
logging.info("Done %d out of %d", data_point_ix,
dataset_size)
logging.info("Training data action counts %r",
action_counts)
num_actions = 0
max_num_actions = len(data_point.get_trajectory())
max_num_actions += self.constants["max_extra_horizon"]
image, metadata = agent.server.reset_receive_feedback(
data_point)
state = AgentObservedState(instruction=data_point.instruction,
config=self.config,
constants=self.constants,
start_image=image,
previous_action=None)
mode = ReadPointerAgent.READ_MODE
last_action_was_halt = False
instruction = instruction_to_string(
data_point.get_instruction(), self.config)
print "TRAIN INSTRUCTION: %r" % instruction
print ""
while True:
# Sample action using the policy
# Generate probabilities over actions
probabilities = list(
torch.exp(self.model.get_probs(state, mode).data))
# Use test policy to get the action
action = gp.sample_action_from_prob(probabilities)
action_counts[mode][action] += 1
if mode == ReadPointerAgent.READ_MODE:
# read mode boundary conditions
forced_action = False
if not state.are_tokens_left_to_be_read():
# force halt
action = 1
forced_action = True
elif num_actions >= max_num_actions or last_action_was_halt:
# force read
action = 0
forced_action = True
if not forced_action:
# Store reward in the replay memory list
reward = self._calc_reward_read_mode(state, action)
replay_item = ReplayMemoryItem(state,
action,
reward,
mode=mode)
batch_replay_items.append(replay_item)
if action == 0:
last_action_was_halt = False
state = state.update_on_read()
elif action == 1:
last_action_was_halt = True
mode = ReadPointerAgent.ACT_MODE
else:
raise AssertionError(
"Read mode only supports two actions: read(0) and halt(1). "
+ "Found " + str(action))
elif mode == ReadPointerAgent.ACT_MODE:
# deal with act mode boundary conditions
if num_actions >= max_num_actions:
forced_stop = True
break
elif action == agent.action_space.get_stop_action_index(
):
if state.are_tokens_left_to_be_read():
reward = self._calc_reward_act_halt(state)
# Add to replay memory
replay_item = ReplayMemoryItem(
state,
agent.action_space.get_stop_action_index(),
reward, mode)
batch_replay_items.append(replay_item)
mode = ReadPointerAgent.READ_MODE
last_action_was_halt = True
state = state.update_on_act_halt()
else:
forced_stop = False
break
else:
image, reward, metadata = agent.server.send_action_receive_feedback(
action)
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state,
action,
reward,
mode=mode)
batch_replay_items.append(replay_item)
# Update the agent state
state = state.update(image, action)
num_actions += 1
total_reward += reward
last_action_was_halt = False
else:
raise AssertionError(
"Mode should be either read or act. Unhandled mode: "
+ str(mode))
assert mode == ReadPointerAgent.ACT_MODE, "Agent should end on Act Mode"
# Send final STOP action and get feedback
image, reward, metadata = agent.server.halt_and_receive_feedback(
)
total_reward += reward
# Store it in the replay memory list
if not forced_stop:
replay_item = ReplayMemoryItem(
state, agent.action_space.get_stop_action_index(),
reward, mode)
batch_replay_items.append(replay_item)
# Update the scores based on meta_data
# self.meta_data_util.log_results(metadata)
# Perform update
episodes_in_batch += 1
if episodes_in_batch == 1:
loss_val = self.do_update(batch_replay_items)
batch_replay_items = []
entropy_val = float(self.entropy.data[0])
self.tensorboard.log(entropy_val, loss_val, total_reward)
total_reward = 0
episodes_in_batch = 0
self.tensorboard.log_train_error(metadata["error"])
# Save the model
self.model.save_model(
experiment_name +
"/read_pointer_contextual_bandit_resnet_epoch_" + str(epoch))
logging.info("Training data action counts %r", action_counts)
def do_train_forced_reading(self, agent, train_dataset, tune_dataset, experiment_name):
""" Perform training """
assert isinstance(agent, ReadPointerAgent), "This learning algorithm works only with READPointerAgent"
dataset_size = len(train_dataset)
for epoch in range(1, self.max_epoch + 1):
logging.info("Starting epoch %d", epoch)
action_counts = dict()
action_counts[ReadPointerAgent.READ_MODE] = [0] * 2
action_counts[ReadPointerAgent.ACT_MODE] = [0] * self.action_space.num_actions()
# Test on tuning data
agent.test_forced_reading(tune_dataset, tensorboard=self.tensorboard)
batch_replay_items = []
total_reward = 0
episodes_in_batch = 0
for data_point_ix, data_point in enumerate(train_dataset):
if (data_point_ix + 1) % 100 == 0:
logging.info("Done %d out of %d", data_point_ix, dataset_size)
logging.info("Training data action counts %r", action_counts)
image, metadata = agent.server.reset_receive_feedback(data_point)
pose = int(metadata["y_angle"] / 15.0)
oracle_segments = data_point.get_instruction_oracle_segmented()
state = AgentObservedState(instruction=data_point.instruction,
config=self.config,
constants=self.constants,
start_image=image,
previous_action=None,
pose=pose)
mode = ReadPointerAgent.READ_MODE
current_segment_ix = 0
trajectories = data_point.get_sub_trajectory_list()
action_ix = 0
while True:
if mode == ReadPointerAgent.READ_MODE:
# Find the number of tokens to read for the gold segment
num_segment_size = len(oracle_segments[current_segment_ix])
current_segment_ix += 1
for i in range(0, num_segment_size):
state = state.update_on_read()
mode = ReadPointerAgent.ACT_MODE
elif mode == ReadPointerAgent.ACT_MODE:
if action_ix == len(trajectories[current_segment_ix - 1]):
action = agent.action_space.get_stop_action_index()
action_ix = 0
else:
action = trajectories[current_segment_ix - 1][action_ix]
action_ix += 1
action_counts[mode][action] += 1
if action == agent.action_space.get_stop_action_index():
if state.are_tokens_left_to_be_read():
# Add to replay memory
replay_item = ReplayMemoryItem(state, agent.action_space.get_stop_action_index(),
1.0, mode)
batch_replay_items.append(replay_item)
mode = ReadPointerAgent.READ_MODE
agent.server.force_goal_update()
state = state.update_on_act_halt()
else:
break
else:
image, reward, metadata = agent.server.send_action_receive_feedback(action)
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state, action, 1, mode=mode)
batch_replay_items.append(replay_item)
# Update the agent state
pose = int(metadata["y_angle"] / 15.0)
state = state.update(image, action, pose=pose)
total_reward += reward
else:
raise AssertionError("Mode should be either read or act. Unhandled mode: " + str(mode))
assert mode == ReadPointerAgent.ACT_MODE, "Agent should end on Act Mode"
# Send final STOP action and get feedback
image, reward, metadata = agent.server.halt_and_receive_feedback()
total_reward += reward
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state, agent.action_space.get_stop_action_index(), 1, mode)
batch_replay_items.append(replay_item)
# Update the scores based on meta_data
# self.meta_data_util.log_results(metadata)
# Perform update
episodes_in_batch += 1
if episodes_in_batch == 1:
loss_val = self.do_update(batch_replay_items)
batch_replay_items = []
entropy_val = float(self.entropy.data[0])
self.tensorboard.log(entropy_val, loss_val, total_reward)
total_reward = 0
episodes_in_batch = 0
self.tensorboard.log_train_error(metadata["error"])
# Save the model
self.model.save_model(
experiment_name + "/ml_estimation_epoch_" + str(epoch))
logging.info("Training data action counts %r", action_counts)
def do_train(self, agent, train_dataset, tune_dataset, experiment_name):
""" Perform training """
dataset_size = len(train_dataset)
for epoch in range(1, self.max_epoch + 1):
logging.info("Starting epoch %d", epoch)
action_counts = [0] * self.action_space.num_actions()
# Test on tuning data
agent.test(tune_dataset, tensorboard=self.tensorboard)
batch_replay_items = []
total_reward = 0
episodes_in_batch = 0
for data_point_ix, data_point in enumerate(train_dataset):
if (data_point_ix + 1) % 100 == 0:
logging.info("Done %d out of %d", data_point_ix,
dataset_size)
logging.info("Training data action counts %r",
action_counts)
# instruction = instruction_to_string(
# data_point.get_instruction(), self.config)
# print "TRAIN INSTRUCTION: %r" % instruction
# print ""
instruction = data_point.get_paragraph_instruction()
num_actions = 0
max_num_actions = len(data_point.get_trajectory())
max_num_actions += self.constants["max_extra_horizon"]
image, metadata = agent.server.reset_receive_feedback(
data_point)
pose = int(metadata["y_angle"] / 15.0)
position_orientation = (metadata["x_pos"], metadata["z_pos"],
metadata["y_angle"])
state = AgentObservedState(
instruction=data_point.get_paragraph_instruction(),
config=self.config,
constants=self.constants,
start_image=image,
previous_action=None,
pose=pose,
position_orientation=position_orientation,
data_point=data_point)
state.start_read_pointer, state.end_read_pointer = data_point.get_instruction_indices(
)
forced_stop = True
while num_actions < max_num_actions:
# Sample action using the policy
# Generate probabilities over actions
probabilities = list(
torch.exp(self.model.get_probs(state).data))
# Use test policy to get the action
action = gp.sample_action_from_prob(probabilities)
action_counts[action] += 1
if action == agent.action_space.get_stop_action_index():
forced_stop = False
break
# Send the action and get feedback
image, reward, metadata = agent.server.send_action_receive_feedback(
action)
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state, action, reward)
batch_replay_items.append(replay_item)
# Update the agent state
pose = int(metadata["y_angle"] / 15.0)
position_orientation = (metadata["x_pos"],
metadata["z_pos"],
metadata["y_angle"])
state = state.update(
image,
action,
pose=pose,
position_orientation=position_orientation,
data_point=data_point)
num_actions += 1
total_reward += reward
# Send final STOP action and get feedback
image, reward, metadata = agent.server.halt_and_receive_feedback(
)
total_reward += reward
# Store it in the replay memory list
if not forced_stop:
replay_item = ReplayMemoryItem(
state, agent.action_space.get_stop_action_index(),
reward)
batch_replay_items.append(replay_item)
# Update the scores based on meta_data
# self.meta_data_util.log_results(metadata)
# Perform update
episodes_in_batch += 1
if episodes_in_batch == 1:
loss_val = self.do_update(batch_replay_items)
batch_replay_items = []
# entropy_val = float(self.entropy.data[0])
# self.tensorboard.log(entropy_val, loss_val, total_reward)
cross_entropy = float(self.cross_entropy.data[0])
self.tensorboard.log(cross_entropy, loss_val, total_reward)
total_reward = 0
episodes_in_batch = 0
if self.tensorboard is not None:
self.tensorboard.log_all_train_errors(
metadata["edit_dist_error"],
metadata["closest_dist_error"],
metadata["stop_dist_error"])
# Save the model
self.model.save_model(experiment_name +
"/contextual_bandit_resnet_epoch_" +
str(epoch))
logging.info("Training data action counts %r", action_counts)
def do_train_(shared_model,
config,
action_space,
meta_data_util,
constants,
train_dataset,
tune_dataset,
experiment,
experiment_name,
rank,
server,
logger,
model_type,
use_pushover=False):
server.initialize_server()
# Test policy
test_policy = gp.get_argmax_action
# torch.manual_seed(args.seed + rank)
if rank == 0: # client 0 creates a tensorboard server
tensorboard = Tensorboard(experiment_name)
else:
tensorboard = None
if use_pushover:
pushover_logger = PushoverLogger(experiment_name)
else:
pushover_logger = None
# Create a local model for rollouts
local_model = model_type(config, constants)
# local_model.train()
# Create the Agent
logger.log("STARTING AGENT")
agent = Agent(server=server,
model=local_model,
test_policy=test_policy,
action_space=action_space,
meta_data_util=meta_data_util,
config=config,
constants=constants)
logger.log("Created Agent...")
action_counts = [0] * action_space.num_actions()
max_epochs = constants["max_epochs"]
dataset_size = len(train_dataset)
tune_dataset_size = len(tune_dataset)
# Create the learner to compute the loss
learner = TmpStreetViewAsynchronousSupervisedLearning(
shared_model, local_model, action_space, meta_data_util, config,
constants, tensorboard)
for epoch in range(1, max_epochs + 1):
learner.epoch = epoch
task_completion_accuracy = 0
mean_stop_dist_error = 0
time_taken = dict()
time_taken["prob_time"] = 0.0
time_taken["update_time"] = 0.0
time_taken["server_time"] = 0.0
time_taken["total_time"] = 0.0
for data_point_ix, data_point in enumerate(train_dataset):
start = time.time()
# Sync with the shared model
# local_model.load_state_dict(shared_model.state_dict())
local_model.load_from_state_dict(shared_model.get_state_dict())
if (data_point_ix + 1) % 100 == 0:
logger.log("Done %d out of %d" %
(data_point_ix, dataset_size))
logger.log("Training data action counts %r" %
action_counts)
logger.log(
"Total Time %f, Server Time %f, Update Time %f, Prob Time %f "
% (time_taken["total_time"], time_taken["server_time"],
time_taken["update_time"], time_taken["prob_time"]))
num_actions = 0
time_start = time.time()
image, metadata = agent.server.reset_receive_feedback(
data_point)
time_taken["server_time"] += time.time() - time_start
state = AgentObservedState(instruction=data_point.instruction,
config=config,
constants=constants,
start_image=image,
previous_action=None,
data_point=data_point)
# state.goal = GoalPrediction.get_goal_location(metadata, data_point,
# learner.image_height, learner.image_width)
model_state = None
batch_replay_items = []
total_reward = 0
trajectory = agent.server.get_trajectory_exact(
data_point.trajectory)
trajectory = trajectory[:min(len(trajectory
), constants["horizon"])]
for action in trajectory:
# Sample action using the policy
time_start = time.time()
log_probabilities, model_state, image_emb_seq, volatile = \
local_model.get_probs(state, model_state)
time_taken["prob_time"] += time.time() - time_start
# Sample action from the probability
action_counts[action] += 1
# Send the action and get feedback
time_start = time.time()
image, reward, metadata = agent.server.send_action_receive_feedback(
action)
time_taken["server_time"] += time.time() - time_start
# Store it in the replay memory list
replay_item = ReplayMemoryItem(state,
action,
reward,
log_prob=log_probabilities,
volatile=volatile,
goal=None)
batch_replay_items.append(replay_item)
# Update the agent state
state = state.update(image, action, data_point=data_point)
# state.goal = GoalPrediction.get_goal_location(metadata, data_point,
# learner.image_height, learner.image_width)
num_actions += 1
total_reward += reward
time_start = time.time()
log_probabilities, model_state, image_emb_seq, volatile = \
local_model.get_probs(state, model_state)
time_taken["prob_time"] += time.time() - time_start
# Send final STOP action and get feedback
time_start = time.time()
image, reward, metadata = agent.server.halt_and_receive_feedback(
)
time_taken["server_time"] += time.time() - time_start
total_reward += reward
if metadata["navigation_error"] <= 5.0:
task_completion_accuracy += 1
mean_stop_dist_error += metadata["navigation_error"]
if tensorboard is not None:
tensorboard.log_scalar("navigation_error",
metadata["navigation_error"])
# Store it in the replay memory list
replay_item = ReplayMemoryItem(
state,
action_space.get_stop_action_index(),
reward,
log_prob=log_probabilities,
volatile=volatile,
goal=None)
batch_replay_items.append(replay_item)
# Update the scores based on meta_data
# self.meta_data_util.log_results(metadata)
# Perform update
time_start = time.time()
if len(batch_replay_items) > 0: # 32:
loss_val = learner.do_update(batch_replay_items)
# self.action_prediction_loss_calculator.predict_action(batch_replay_items)
# del batch_replay_items[:] # in place list clear
if tensorboard is not None:
cross_entropy = 0.0 # float(learner.cross_entropy.data[0])
tensorboard.log(cross_entropy, loss_val, 0)
entropy = float(
learner.entropy.data[0]) / float(num_actions + 1)
logger.log(
"Entropy %r, Total Reward %r, Loss %r, Num Actions %d, stop-error %r "
% (entropy, total_reward, loss_val,
num_actions + 1, metadata["navigation_error"]))
tensorboard.log_scalar("entropy", entropy)
tensorboard.log_scalar("total_reward", total_reward)
time_taken["update_time"] += time.time() - time_start
time_taken["total_time"] += time.time() - start
# Save the model
local_model.save_model(experiment + "/supervised_learning" +
str(rank) + "_epoch_" + str(epoch))
logger.log("Training data action counts %r" % action_counts)
mean_stop_dist_error = mean_stop_dist_error / float(
len(train_dataset))
task_completion_accuracy = (task_completion_accuracy *
100.0) / float(len(train_dataset))
logger.log("Training: Mean stop distance error %r" %
mean_stop_dist_error)
logger.log("Training: Task completion accuracy %r " %
task_completion_accuracy)
if tune_dataset_size > 0:
# Test on tuning data
agent.test(tune_dataset,
tensorboard=tensorboard,
logger=logger,
pushover_logger=pushover_logger)