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(chaplot_baseline,
shared_model,
config,
action_space,
meta_data_util,
args,
constants,
train_dataset,
tune_dataset,
experiment,
experiment_name,
rank,
server,
logger,
model_type,
contextual_bandit,
use_pushover=False):
sys.stderr = sys.stdout
server.initialize_server()
# Local Config Variables
lstm_size = 256
# 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
# Create the Agent
logger.log("STARTING AGENT")
agent = Agent(server=server,
model=chaplot_baseline,
test_policy=test_policy,
action_space=action_space,
meta_data_util=meta_data_util,
config=config,
constants=constants)
logger.log("Created Agent...")
# Create a local model for rollouts
local_model = model_type(args, config=config)
if torch.cuda.is_available():
local_model.cuda()
chaplot_baseline.shared_model = local_model
local_model.train()
# Our Environment Interface
env = NavDroneServerInterface(agent, local_model, experiment, config,
constants, None, train_dataset,
tune_dataset, rank, logger, use_pushover)
env.game_init()
# logging.info("Contextual bandit is %r and horizon is %r", self.contextual_bandit, args.max_episode_length)
logger.log("Created NavDroneServerInterface")
# optimizer = optim.SGD(self.shared_model.parameters(), lr=self.args.lr) --- changed Chaplot's optimizer
optimizer = optim.Adam(shared_model.parameters(), lr=0.00025)
p_losses = []
v_losses = []
launch_k_unity_builds([config["port"]],
"./simulators/NavDroneLinuxBuild.x86_64")
(image, instr), _, _ = env.reset()
curr_instr, prev_instr, next_instr = instr
curr_instruction_idx = np.array(curr_instr)
prev_instruction_idx = np.array(prev_instr)
next_instruction_idx = np.array(next_instr)
image = torch.from_numpy(image).float()
curr_instruction_idx = torch.from_numpy(curr_instruction_idx).view(
1, -1)
prev_instruction_idx = torch.from_numpy(prev_instruction_idx).view(
1, -1)
next_instruction_idx = torch.from_numpy(next_instruction_idx).view(
1, -1)
done = True
episode_length = 0
num_iters = 0
while True:
# Sync with the shared model
local_model.load_state_dict(shared_model.state_dict())
if done:
episode_length = 0
cx = Variable(torch.zeros(1, lstm_size).cuda())
hx = Variable(torch.zeros(1, lstm_size).cuda())
else:
# assert False, "Assertion put by Max and Dipendra. Code shouldn't reach here."
cx = Variable(cx.data.cuda())
hx = Variable(hx.data.cuda())
values = []
log_probs = []
rewards = []
entropies = []
cached_information = None
for step in range(args.num_steps):
episode_length += 1
tx = Variable(
torch.from_numpy(np.array([episode_length])).long().cuda())
value, logit, (hx, cx), cached_information = local_model(
(Variable(image.unsqueeze(0).cuda()),
Variable(curr_instruction_idx.cuda()),
Variable(prev_instruction_idx.cuda()),
Variable(next_instruction_idx.cuda()), (tx, hx, cx)),
cached_information)
prob = F.softmax(logit, dim=1)
log_prob = F.log_softmax(logit, dim=1)
entropy = -(log_prob * prob).sum(1)
entropies.append(entropy)
action = prob.multinomial().data
log_prob = log_prob.gather(1, Variable(action.cuda()))
action = action.cpu().numpy()[0, 0]
(image, _), reward, done, _ = env.step(action)
# done = done or (episode_length >= self.args.max_episode_length)
if not done and (episode_length >= args.max_episode_length):
# If the agent has not taken
_, _, done, _ = env.step(
env.client.agent.action_space.get_stop_action_index())
done = True
if done:
(image, instr), _, _ = env.reset()
curr_instr, prev_instr, next_instr = instr
curr_instruction_idx = np.array(curr_instr)
prev_instruction_idx = np.array(prev_instr)
next_instruction_idx = np.array(next_instr)
curr_instruction_idx = torch.from_numpy(
curr_instruction_idx).view(1, -1)
prev_instruction_idx = torch.from_numpy(
prev_instruction_idx).view(1, -1)
next_instruction_idx = torch.from_numpy(
next_instruction_idx).view(1, -1)
image = torch.from_numpy(image).float()
values.append(value)
log_probs.append(log_prob)
rewards.append(reward)
if done:
break
if rank == 0 and tensorboard is not None:
# Log total reward and entropy
tensorboard.log_scalar("Total_Reward", sum(rewards))
mean_entropy = sum(entropies).data[0] / float(
max(episode_length, 1))
tensorboard.log_scalar("Chaplot_Baseline_Entropy",
mean_entropy)
R = torch.zeros(1, 1)
if not done:
tx = Variable(
torch.from_numpy(np.array([episode_length])).long().cuda())
value, _, _, _ = local_model(
(Variable(image.unsqueeze(0).cuda()),
Variable(curr_instruction_idx.cuda()),
Variable(prev_instruction_idx.cuda()),
Variable(next_instruction_idx.cuda()), (tx, hx, cx)))
R = value.data
values.append(Variable(R.cuda()))
policy_loss = 0
value_loss = 0
R = Variable(R.cuda())
gae = torch.zeros(1, 1).cuda()
for i in reversed(range(len(rewards))):
R = args.gamma * R + rewards[i]
advantage = R - values[i]
value_loss = value_loss + 0.5 * advantage.pow(2)
if contextual_bandit:
# Just focus on immediate reward
gae = torch.from_numpy(np.array([[rewards[i]]])).float()
else:
# Generalized Advantage Estimataion
delta_t = rewards[i] + args.gamma * \
values[i + 1].data - values[i].data
gae = gae * args.gamma * args.tau + delta_t
policy_loss = policy_loss - \
log_probs[i] * Variable(gae.cuda()) - 0.02 * entropies[i]
optimizer.zero_grad()
p_losses.append(policy_loss.data[0, 0])
v_losses.append(value_loss.data[0, 0])
if len(p_losses) > 1000:
num_iters += 1
logger.log(" ".join([
# "Training thread: {}".format(rank),
"Num iters: {}K".format(num_iters),
"Avg policy loss: {}".format(np.mean(p_losses)),
"Avg value loss: {}".format(np.mean(v_losses))
]))
p_losses = []
v_losses = []
(policy_loss + 0.5 * value_loss).backward()
torch.nn.utils.clip_grad_norm(local_model.parameters(), 40)
ChaplotBaseline.ensure_shared_grads(local_model, shared_model)
optimizer.step()
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_(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_(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_supervsed_train(chaplot_baseline,
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):
raise NotImplementedError()
sys.stderr = sys.stdout
server.initialize_server()
# Local Config Variables
lstm_size = 256
# 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
# Create the Agent
logger.log("STARTING AGENT")
agent = Agent(server=server,
model=chaplot_baseline,
test_policy=test_policy,
action_space=action_space,
meta_data_util=meta_data_util,
config=config,
constants=constants)
logger.log("Created Agent...")
# Create a local model for rollouts
local_model = model_type(args, config=config)
if torch.cuda.is_available():
local_model.cuda()
chaplot_baseline.shared_model = local_model
local_model.train()
env = StreetViewServerInterface(agent, local_model, experiment, config,
constants, None, train_dataset,
tune_dataset, rank, logger,
use_pushover)
env.game_init()
shared_model.train()
# optimizer = optim.SGD(self.shared_model.parameters(), lr=self.args.lr)
optimizer = optim.Adam(shared_model.parameters(), lr=0.00025)
p_losses = []
v_losses = []
num_iters = 0
while True:
# Get datapoint
(image, instr), _, _ = env.reset()
instruction_idx = np.array(instr)
image = torch.from_numpy(image).float()
instruction_idx = torch.from_numpy(instruction_idx).view(1, -1)
# Sync with the shared model
# model.load_state_dict(shared_model.state_dict())
episode_length = 0
cx = Variable(torch.zeros(1, lstm_size).cuda())
hx = Variable(torch.zeros(1, lstm_size).cuda())
log_probs = []
rewards = []
entropies = []
trajectory = env.get_trajectory()
min_length = min(len(trajectory), args.max_episode_length - 1)
trajectory = trajectory[0:min_length]
trajectory.append(agent.action_space.get_stop_action_index())
for action in trajectory:
episode_length += 1
tx = Variable(
torch.from_numpy(np.array([episode_length])).long().cuda())
value, logit, (hx, cx) = shared_model(
(Variable(image.unsqueeze(0).cuda()),
Variable(instruction_idx.cuda()), None, None, (tx, hx,
cx)))
prob = F.softmax(logit)
log_prob = F.log_softmax(logit)
entropy = -(log_prob * prob).sum(1)
entropies.append(entropy)
action_tensor = torch.from_numpy(np.array([[action]]))
log_prob = log_prob.gather(1, Variable(action_tensor.cuda()))
(image, _), reward, done, _ = env.step(action)
image = torch.from_numpy(image).float()
log_probs.append(log_prob)
rewards.append(reward)
if done:
break
policy_loss = 0
for i in range(0, len(rewards)):
policy_loss = policy_loss - log_probs[i] - 0.01 * entropies[i]
# Log total reward and entropy
if tensorboard is not None:
tensorboard.log_scalar("Total_Reward", sum(rewards))
mean_entropy = sum(entropies) / float(max(episode_length, 1))
tensorboard.log_scalar("Chaplot_Baseline_Entropy",
mean_entropy)
tensorboard.log_scalar("Policy_Loss", policy_loss)
optimizer.zero_grad()
p_losses.append(policy_loss.data[0, 0])
if len(p_losses) > 1000:
num_iters += 1
logger.log(" ".join([
# "Training thread: {}".format(rank),
"Num iters: {}K".format(num_iters),
"Avg policy loss: {}".format(np.mean(p_losses)),
"Avg value loss: {}".format(np.mean(v_losses))
]))
p_losses = []
v_losses = []
policy_loss.backward()
torch.nn.utils.clip_grad_norm(shared_model.parameters(), 40)
# ensure_shared_grads(model, shared_model)
optimizer.step()
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_(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_supervised_train(chaplot_baseline,
shared_model,
config,
action_space,
meta_data_util,
args,
constants,
train_dataset,
tune_dataset,
experiment,
experiment_name,
rank,
server,
logger,
model_type,
contextual_bandit=False,
use_pushover=False):
try:
sys.stderr = sys.stdout
server.initialize_server()
# Local Config Variables
lstm_size = 256
# 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
# Create the Agent
logger.log("STARTING AGENT")
agent = Agent(server=server,
model=chaplot_baseline,
test_policy=test_policy,
action_space=action_space,
meta_data_util=meta_data_util,
config=config,
constants=constants)
logger.log("Created Agent...")
# Create a local model for rollouts
local_model = model_type(args, config=config)
if torch.cuda.is_available():
local_model.cuda()
chaplot_baseline.shared_model = local_model
local_model.train()
# Our Environment Interface
env = NavDroneServerInterface(agent, local_model, experiment,
config, constants, None,
train_dataset, tune_dataset, rank,
logger, use_pushover)
logger.log("Created NavDroneServerInterface")
# optimizer = optim.SGD(self.shared_model.parameters(), lr=self.args.lr) --- changed Chaplot's optimizer
optimizer = optim.Adam(shared_model.parameters(), lr=0.00025)
p_losses = []
v_losses = []
launch_k_unity_builds([
config["port"]
], "/home/dipendra/Downloads/NavDroneLinuxBuild/NavDroneLinuxBuild.x86_64"
)
done = True
num_iters = 0
global_id = 1
while True:
# Sync with the shared model
local_model.load_state_dict(shared_model.state_dict())
# Get a new datapoint
(image, instr), _, _, metadata, data_point = env.reset()
curr_instr, prev_instr, next_instr = instr
curr_instruction_idx = np.array(curr_instr)
prev_instruction_idx = np.array(prev_instr)
next_instruction_idx = np.array(next_instr)
image = torch.from_numpy(image).float()
curr_instruction_idx = torch.from_numpy(
curr_instruction_idx).view(1, -1)
prev_instruction_idx = torch.from_numpy(
prev_instruction_idx).view(1, -1)
next_instruction_idx = torch.from_numpy(
next_instruction_idx).view(1, -1)
episode_length = 0
cx = Variable(torch.zeros(1, lstm_size).cuda())
hx = Variable(torch.zeros(1, lstm_size).cuda())
goal_x, goal_z = data_point.get_destination_list()[-1]
trajectory_str = get_oracle_trajectory(metadata, goal_x,
goal_z, data_point)
trajectory = [
action_space.get_action_index(act_str)
for act_str in trajectory_str
]
# trajectory = data_point.get_trajectory()
num_steps = len(trajectory) + 1 # 1 for stopping
values = []
log_probs = []
rewards = []
entropies = []
cached_information = None
#############################
lstm_rep = []
image_rep = []
actions = []
goal_locations = []
#############################
for step in range(num_steps):
episode_length += 1
tx = Variable(
torch.from_numpy(np.array([episode_length
])).long().cuda())
value, logit, (hx, cx), cached_information = local_model(
(Variable(image.unsqueeze(0).cuda()),
Variable(curr_instruction_idx.cuda()),
Variable(prev_instruction_idx.cuda()),
Variable(next_instruction_idx.cuda()), (tx, hx, cx)),
cached_information)
prob = F.softmax(logit, dim=1)
log_prob = F.log_softmax(logit, dim=1)
entropy = -(log_prob * prob).sum(1)
entropies.append(entropy)
if step == len(trajectory):
action = action_space.get_stop_action_index()
else:
action = trajectory[step]
action_var = torch.from_numpy(np.array([[action]]))
####################################
lstm_rep.append(cached_information["lstm_rep"])
image_rep.append(cached_information["image_rep"])
actions.append(action_var)
goal_location = ChaplotBaselineWithAuxiliary.get_goal_location(
metadata, data_point)
goal_locations.append(goal_location)
# ChaplotBaselineWithAuxiliary.save_visualized_image(image, goal_location, global_id)
global_id += 1
####################################
log_prob = log_prob.gather(1, Variable(action_var.cuda()))
(image, _), reward, done, _, metadata = env.step(action)
image = torch.from_numpy(image).float()
values.append(value)
log_probs.append(log_prob)
rewards.append(reward)
assert done, "Should be done as all trajectories are fully executed and stop with 'stop' action."
if rank == 0 and tensorboard is not None:
# Log total reward and entropy
tensorboard.log_scalar("Total_Reward", sum(rewards))
mean_entropy = sum(entropies).data[0] / float(
max(episode_length, 1))
tensorboard.log_scalar("Chaplot_Baseline_Entropy",
mean_entropy)
R = torch.zeros(1, 1)
values.append(Variable(R.cuda()))
policy_loss = 0
value_loss = 0
R = Variable(R.cuda())
entropy_coeff = max(0.0, 0.11 - env.num_epochs * 0.01)
for i in reversed(range(len(rewards))):
R = args.gamma * R + rewards[i]
advantage = R - values[i]
value_loss = value_loss + 0.5 * advantage.pow(2)
policy_loss = policy_loss - \
log_probs[i] - entropy_coeff * entropies[i]
temporal_autoencoding_loss = None # local_model.get_tae_loss(image_rep, actions)
reward_prediction_loss = None # local_model.get_reward_prediction_loss(lstm_rep, actions, rewards)
alignment_loss, alignment_norm = None, None # local_model.alignment_auxiliary(image_rep, cached_information["text_rep"])
goal_prediction_loss = local_model.calc_goal_prediction_loss(
image_rep, cached_information["text_rep"], goal_locations)
optimizer.zero_grad()
p_losses.append(policy_loss.data[0, 0])
v_losses.append(value_loss.data[0, 0])
if len(p_losses) > 1000:
num_iters += 1
logger.log(" ".join([
# "Training thread: {}".format(rank),
"Num iters: {}K".format(num_iters),
"Avg policy loss: {}".format(np.mean(p_losses)),
"Avg value loss: {}".format(np.mean(v_losses))
]))
p_losses = []
v_losses = []
if rank == 0 and tensorboard is not None:
# Log total reward and entropy
tensorboard.log_scalar("Value_Loss",
float(value_loss.data))
if temporal_autoencoding_loss is not None:
tensorboard.log_scalar(
"TAE_Loss", float(temporal_autoencoding_loss.data))
if reward_prediction_loss is not None:
tensorboard.log_scalar(
"RP_Loss", float(reward_prediction_loss.data))
if alignment_loss is not None:
tensorboard.log_scalar(
"Mean_Current_Segment_Alignment_Loss",
float(alignment_loss.data))
tensorboard.log_scalar("Alignment_Norm",
float(alignment_norm.data))
if goal_prediction_loss is not None:
tensorboard.log_scalar(
"Goal_Prediction_Loss",
float(goal_prediction_loss.data) /
float(num_steps))
loss = policy_loss + 0.5 * value_loss
if temporal_autoencoding_loss is not None:
loss += 0.5 * temporal_autoencoding_loss
if reward_prediction_loss is not None:
loss += 0.5 * reward_prediction_loss
if alignment_loss is not None:
loss += 0.5 * alignment_loss
if goal_prediction_loss is not None:
loss += 20.0 * goal_prediction_loss
loss = goal_prediction_loss
loss.backward()
torch.nn.utils.clip_grad_norm(local_model.parameters(), 40)
ChaplotBaselineWithAuxiliary.ensure_shared_grads(
local_model, shared_model)
optimizer.step()
except Exception:
print("Exception")
exc_info = sys.exc_info()
traceback.print_exception(*exc_info)
def do_train(chaplot_baseline,
shared_model,
config,
action_space,
meta_data_util,
args,
constants,
train_dataset,
tune_dataset,
experiment,
experiment_name,
rank,
server,
logger,
model_type,
contextual_bandit=False,
use_pushover=False):
try:
sys.stderr = sys.stdout
server.initialize_server()
# Local Config Variables
lstm_size = 256
# 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
# Create the Agent
logger.log("STARTING AGENT")
agent = Agent(server=server,
model=chaplot_baseline,
test_policy=test_policy,
action_space=action_space,
meta_data_util=meta_data_util,
config=config,
constants=constants)
logger.log("Created Agent...")
# Create a local model for rollouts
local_model = model_type(args, config=config)
if torch.cuda.is_available():
local_model.cuda()
chaplot_baseline.shared_model = local_model
local_model.train()
# Our Environment Interface
env = NavDroneServerInterface(agent, local_model, experiment,
config, constants, None,
train_dataset, tune_dataset, rank,
logger, use_pushover)
logger.log("Created NavDroneServerInterface")
# optimizer = optim.SGD(self.shared_model.parameters(), lr=self.args.lr) --- changed Chaplot's optimizer
optimizer = optim.Adam(shared_model.parameters(), lr=0.00025)
p_losses = []
v_losses = []
launch_k_unity_builds([config["port"]],
"./simulators/NavDroneLinuxBuild.x86_64")
(image, instr), _, _, metadata, data_point = env.reset()
curr_instr, prev_instr, next_instr = instr
curr_instruction_idx = np.array(curr_instr)
prev_instruction_idx = np.array(prev_instr)
next_instruction_idx = np.array(next_instr)
image = torch.from_numpy(image).float()
curr_instruction_idx = torch.from_numpy(curr_instruction_idx).view(
1, -1)
prev_instruction_idx = torch.from_numpy(prev_instruction_idx).view(
1, -1)
next_instruction_idx = torch.from_numpy(next_instruction_idx).view(
1, -1)
done = True
episode_length = 0
num_iters = 0
cx, hx = None, None
global_id = 1
while True:
# Sync with the shared model
local_model.load_state_dict(shared_model.state_dict())
if done:
episode_length = 0
cx = Variable(torch.zeros(1, lstm_size).cuda())
hx = Variable(torch.zeros(1, lstm_size).cuda())
else:
cx = Variable(cx.data.cuda())
hx = Variable(hx.data.cuda())
values = []
log_probs = []
rewards = []
entropies = []
cached_information = None
#############################
lstm_rep = []
image_rep = []
actions = []
goal_locations = []
#############################
for step in range(args.num_steps):
episode_length += 1
tx = Variable(
torch.from_numpy(np.array([episode_length
])).long().cuda())
value, logit, (hx, cx), cached_information = local_model(
(Variable(image.unsqueeze(0).cuda()),
Variable(curr_instruction_idx.cuda()),
Variable(prev_instruction_idx.cuda()),
Variable(next_instruction_idx.cuda()), (tx, hx, cx)),
cached_information)
prob = F.softmax(logit, dim=1)
log_prob = F.log_softmax(logit, dim=1)
entropy = -(log_prob * prob).sum(1)
entropies.append(entropy)
action = prob.multinomial().data
####################################
lstm_rep.append(cached_information["lstm_rep"])
image_rep.append(cached_information["image_rep"])
actions.append(action)
goal_location = ChaplotBaselineWithAuxiliary.get_goal_location(
metadata, data_point)
goal_locations.append(goal_location)
# ChaplotBaselineWithAuxiliary.save_visualized_image(image, goal_location, global_id)
global_id += 1
####################################
log_prob = log_prob.gather(1, Variable(action.cuda()))
action = action.cpu().numpy()[0, 0]
(image, _), reward, done, _, metadata = env.step(action)
# done = done or (episode_length >= self.args.max_episode_length)
if not done and (episode_length >=
args.max_episode_length):
# If the agent has not taken
_, _, done, _, metadata = env.step(
env.agent.action_space.get_stop_action_index())
done = True
if done:
(image,
instr), _, _, metadata, data_point = env.reset()
curr_instr, prev_instr, next_instr = instr
curr_instruction_idx = np.array(curr_instr)
prev_instruction_idx = np.array(prev_instr)
next_instruction_idx = np.array(next_instr)
curr_instruction_idx = torch.from_numpy(
curr_instruction_idx).view(1, -1)
prev_instruction_idx = torch.from_numpy(
prev_instruction_idx).view(1, -1)
next_instruction_idx = torch.from_numpy(
next_instruction_idx).view(1, -1)
image = torch.from_numpy(image).float()
values.append(value)
log_probs.append(log_prob)
rewards.append(reward)
if done:
break
if rank == 0 and tensorboard is not None:
# Log total reward and entropy
tensorboard.log_scalar("Total_Reward", sum(rewards))
mean_entropy = sum(entropies).data[0] / float(
max(episode_length, 1))
tensorboard.log_scalar("Chaplot_Baseline_Entropy",
mean_entropy)
R = torch.zeros(1, 1)
if not done:
tx = Variable(
torch.from_numpy(np.array([episode_length
])).long().cuda())
value, _, _, _ = local_model(
(Variable(image.unsqueeze(0).cuda()),
Variable(curr_instruction_idx.cuda()),
Variable(prev_instruction_idx.cuda()),
Variable(next_instruction_idx.cuda()), (tx, hx, cx)))
R = value.data
values.append(Variable(R.cuda()))
policy_loss = 0
value_loss = 0
R = Variable(R.cuda())
gae = torch.zeros(1, 1).cuda()
entropy_coeff = max(0.0, 0.11 - env.num_epochs * 0.01)
for i in reversed(range(len(rewards))):
R = args.gamma * R + rewards[i]
advantage = R - values[i]
value_loss = value_loss + 0.5 * advantage.pow(2)
if contextual_bandit:
# Just focus on immediate reward
gae = torch.from_numpy(np.array([[rewards[i]]
])).float()
else:
# Generalized Advantage Estimataion
delta_t = rewards[i] + args.gamma * \
values[i + 1].data - values[i].data
gae = gae * args.gamma * args.tau + delta_t
policy_loss = policy_loss - \
log_probs[i] * Variable(gae.cuda()) - entropy_coeff * entropies[i]
temporal_autoencoding_loss = None # local_model.get_tae_loss(image_rep, actions)
reward_prediction_loss = None # local_model.get_reward_prediction_loss(lstm_rep, actions, rewards)
alignment_loss, alignment_norm = None, None # local_model.alignment_auxiliary(image_rep, cached_information["text_rep"])
goal_prediction_loss = local_model.calc_goal_prediction_loss(
image_rep, cached_information["text_rep"], goal_locations)
optimizer.zero_grad()
p_losses.append(policy_loss.data[0, 0])
v_losses.append(value_loss.data[0, 0])
if len(p_losses) > 1000:
num_iters += 1
logger.log(" ".join([
# "Training thread: {}".format(rank),
"Num iters: {}K".format(num_iters),
"Avg policy loss: {}".format(np.mean(p_losses)),
"Avg value loss: {}".format(np.mean(v_losses))
]))
p_losses = []
v_losses = []
if rank == 0 and tensorboard is not None:
if done:
tensorboard.log_scalar("train_dist_error",
metadata["stop_dist_error"])
task_completion = 0
if metadata["stop_dist_error"] < 5.0:
task_completion = 1
tensorboard.log_scalar("train_task_completion",
task_completion)
# Log total reward and entropy
tensorboard.log_scalar("Value_Loss",
float(value_loss.data))
if temporal_autoencoding_loss is not None:
tensorboard.log_scalar(
"TAE_Loss", float(temporal_autoencoding_loss.data))
if reward_prediction_loss is not None:
tensorboard.log_scalar(
"RP_Loss", float(reward_prediction_loss.data))
if alignment_loss is not None:
tensorboard.log_scalar(
"Mean_Current_Segment_Alignment_Loss",
float(alignment_loss.data))
tensorboard.log_scalar("Alignment_Norm",
float(alignment_norm.data))
if goal_prediction_loss is not None:
tensorboard.log_scalar(
"Goal_Prediction_Loss",
float(goal_prediction_loss.data) /
float(len(rewards)))
loss = policy_loss + 0.5 * value_loss
if temporal_autoencoding_loss is not None:
loss += 0.5 * temporal_autoencoding_loss
if reward_prediction_loss is not None:
loss += 0.5 * reward_prediction_loss
if alignment_loss is not None:
loss += 0.5 * alignment_loss
if goal_prediction_loss is not None:
loss += 0.5 * goal_prediction_loss
loss.backward()
torch.nn.utils.clip_grad_norm(local_model.parameters(), 40)
ChaplotBaselineWithAuxiliary.ensure_shared_grads(
local_model, shared_model)
optimizer.step()
except Exception:
print("Exception")
exc_info = sys.exc_info()
traceback.print_exception(*exc_info)
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)