def run(learning_rate, freeze_interval, num_hidden, reg):
room_size = 5
num_rooms = 2
mdp = mdps.MazeMDP(room_size, num_rooms)
mdp.compute_states()
mdp.EXIT_REWARD = 1
mdp.MOVE_REWARD = -0.01
discount = 1
num_actions = len(mdp.get_actions(None))
batch_size = 100
print 'building network...'
network = qnetwork.QNetwork(input_shape=2 * room_size + num_rooms ** 2, batch_size=batch_size, num_hidden_layers=2, num_actions=4, num_hidden=num_hidden, discount=discount, learning_rate=learning_rate, regularization=reg, update_rule='adam', freeze_interval=freeze_interval, rng=None)
num_epochs = 50
epoch_length = 2
test_epoch_length = 0
max_steps = 4 * (room_size * num_rooms) ** 2
epsilon_decay = (num_epochs * epoch_length * max_steps) / 1.5
print 'building policy...'
p = policy.EpsilonGreedy(num_actions, 0.5, 0.05, epsilon_decay)
print 'building memory...'
rm = replay_memory.ReplayMemory(batch_size, capacity=50000)
print 'building logger...'
log = logger.NeuralLogger(agent_name='QNetwork')
print 'building state adapter...'
adapter = state_adapters.CoordinatesToRowColRoomAdapter(room_size=room_size, num_rooms=num_rooms)
# adapter = state_adapters.CoordinatesToRowColAdapter(room_size=room_size, num_rooms=num_rooms)
# adapter = state_adapters.CoordinatesToFlattenedGridAdapter(room_size=room_size, num_rooms=num_rooms)
# adapter = state_adapters.IdentityAdapter(room_size=room_size, num_rooms=num_rooms)
# adapter = state_adapters.CoordinatesToSingleRoomRowColAdapter(room_size=room_size)
print 'building agent...'
a = agent.NeuralAgent(network=network, policy=p, replay_memory=rm, log=log, state_adapter=adapter)
run_tests = False
e = experiment.Experiment(mdp, a, num_epochs, epoch_length, test_epoch_length, max_steps, run_tests, value_logging=True)
e.run()
ak = file_utils.load_key('../access_key.key')
sk = file_utils.load_key('../secret_key.key')
bucket = 'hierarchical'
try:
aws_util = aws_s3_utility.S3Utility(ak, sk, bucket)
aws_util.upload_directory(e.agent.logger.log_dir)
except Exception as e:
print 'error uploading to s3: {}'.format(e)
# radius of the Earth
R = 6378.1
# radius of images around center of city
IMAGE_RADIUS = 10
# number of images to download from each city
NUM_IMAGES_PER_CITY = 200
# size of failed-download image
FAILED_DOWNLOAD_IMAGE_SIZE = 3554
# place key in a file in the Geo-Localization directory
# as the only text in the file on one line
KEY_FILEPATH = "../api_key.key"
API_KEY = file_utils.load_key(KEY_FILEPATH)
GOOGLE_URL = ("http://maps.googleapis.com/maps/api/streetview?sensor=false&"
"size=256x256&fov=120&key=" + API_KEY)
IMAGES_DIR = '../imgs/'
def download_images_for_city(city, lat, lon):
print 'downloading images of {}'.format(city)
num_imgs = 0
misses = 0
cur_directory = os.path.join(IMAGES_DIR, city)
if not os.path.exists(cur_directory):
os.makedirs(cur_directory)
while num_imgs < 100:
# radius of the Earth
R = 6378.1
# radius of images around center of city
IMAGE_RADIUS = 10
# number of images to download from each city
NUM_IMAGES_PER_CITY = 200
# size of failed-download image
FAILED_DOWNLOAD_IMAGE_SIZE = 3464
# place key in a file in the Geo-Localization directory
# as the only text in the file on one line
KEY_FILEPATH = "../api_key.key"
API_KEY = file_utils.load_key(KEY_FILEPATH)
GOOGLE_URL = ("http://maps.googleapis.com/maps/api/streetview?"
"size=256x256&fov=120&pitch=10&key=" + API_KEY)
IMAGES_DIR = '../data/cities/'
def download_images_for_city(city, lat, lon):
print('downloading images of {}'.format(city))
num_imgs = 0
misses = 0
cur_directory = os.path.join(IMAGES_DIR, city)
if not os.path.exists(cur_directory):
os.makedirs(cur_directory)
while num_imgs < 100:
def run(learning_rate, freeze_interval, num_hidden, reg, seq_len, eps, nt, update):
room_size = 5
num_rooms = 2
input_shape = 2 * room_size
print 'building mdp...'
mdp = mdps.MazeMDP(room_size, num_rooms)
mdp.compute_states()
mdp.EXIT_REWARD = 1
mdp.MOVE_REWARD = -0.01
network_type = nt
discount = 1
sequence_length = seq_len
num_actions = len(mdp.get_actions(None))
batch_size = 100
update_rule = update
print 'building network...'
network = recurrent_qnetwork.RecurrentQNetwork(input_shape=input_shape,
sequence_length=sequence_length, batch_size=batch_size,
num_actions=4, num_hidden=num_hidden, discount=discount,
learning_rate=learning_rate, regularization=reg,
update_rule=update_rule, freeze_interval=freeze_interval,
network_type=network_type, rng=None)
# take this many steps because (very loosely):
# let l be the step length
# let d be the difference in start and end locations
# let N be the number of steps for the agent to travel a distance d
# then N ~ (d/l)^2 // assuming this is a random walk
# with l = 1, this gives d^2 in order to make it N steps away
# the desired distance here is to walk along both dimensions of the maze
# which is equal to two times the num_rooms * room_size
# so squaring that gives a loose approximation to the number of
# steps needed (discounting that this is actually a lattice (does it really matter?))
# (also discounting the walls)
# see: http://mathworld.wolfram.com/RandomWalk2-Dimensional.html
max_steps = (2 * room_size * num_rooms) ** 2
num_epochs = 500
epoch_length = 1
test_epoch_length = 0
epsilon_decay = (num_epochs * epoch_length * max_steps) / 4
print 'building adapter...'
adapter = state_adapters.CoordinatesToSingleRoomRowColAdapter(room_size=room_size)
print 'building policy...'
p = policy.EpsilonGreedy(num_actions, eps, 0.05, epsilon_decay)
print 'building replay memory...'
# want to track at minimum the last 50 episodes
capacity = max_steps * 50
rm = replay_memory.SequenceReplayMemory(input_shape=input_shape,
sequence_length=sequence_length, batch_size=batch_size, capacity=capacity)
print 'building logger...'
log = logger.NeuralLogger(agent_name=network_type)
print 'building agent...'
a = agent.RecurrentNeuralAgent(network=network, policy=p,
replay_memory=rm, log=log, state_adapter=adapter)
run_tests = False
print 'building experiment...'
e = experiment.Experiment(mdp, a, num_epochs, epoch_length,
test_epoch_length, max_steps, run_tests, value_logging=True)
print 'running experiment...'
e.run()
ak = file_utils.load_key('../access_key.key')
sk = file_utils.load_key('../secret_key.key')
bucket = 'hierarchical9'
try:
aws_util = aws_s3_utility.S3Utility(ak, sk, bucket)
aws_util.upload_directory(e.agent.logger.log_dir)
except Exception as e:
print 'error uploading to s3: {}'.format(e)
def run(learning_rate, freeze_interval, num_hidden, reg, seq_len, eps,
nt, update):
room_size = 5
num_rooms = 2
input_shape = 2 * room_size
print 'building mdp...'
mdp = mdps.MazeMDP(room_size, num_rooms)
mdp.compute_states()
mdp.EXIT_REWARD = 1
mdp.MOVE_REWARD = -0.01
network_type = nt
discount = 1
sequence_length = seq_len
num_actions = len(mdp.get_actions(None))
batch_size = 100
update_rule = update
print 'building network...'
network = recurrent_qnetwork.RecurrentQNetwork(
input_shape=input_shape,
sequence_length=sequence_length,
batch_size=batch_size,
num_actions=4,
num_hidden=num_hidden,
discount=discount,
learning_rate=learning_rate,
regularization=reg,
update_rule=update_rule,
freeze_interval=freeze_interval,
network_type=network_type,
rng=None)
# take this many steps because (very loosely):
# let l be the step length
# let d be the difference in start and end locations
# let N be the number of steps for the agent to travel a distance d
# then N ~ (d/l)^2 // assuming this is a random walk
# with l = 1, this gives d^2 in order to make it N steps away
# the desired distance here is to walk along both dimensions of the maze
# which is equal to two times the num_rooms * room_size
# so squaring that gives a loose approximation to the number of
# steps needed (discounting that this is actually a lattice (does it really matter?))
# (also discounting the walls)
# see: http://mathworld.wolfram.com/RandomWalk2-Dimensional.html
max_steps = (2 * room_size * num_rooms)**2
num_epochs = 500
epoch_length = 1
test_epoch_length = 0
epsilon_decay = (num_epochs * epoch_length * max_steps) / 4
print 'building adapter...'
adapter = state_adapters.CoordinatesToSingleRoomRowColAdapter(
room_size=room_size)
print 'building policy...'
p = policy.EpsilonGreedy(num_actions, eps, 0.05, epsilon_decay)
print 'building replay memory...'
# want to track at minimum the last 50 episodes
capacity = max_steps * 50
rm = replay_memory.SequenceReplayMemory(
input_shape=input_shape,
sequence_length=sequence_length,
batch_size=batch_size,
capacity=capacity)
print 'building logger...'
log = logger.NeuralLogger(agent_name=network_type)
print 'building agent...'
a = agent.RecurrentNeuralAgent(network=network,
policy=p,
replay_memory=rm,
log=log,
state_adapter=adapter)
run_tests = False
print 'building experiment...'
e = experiment.Experiment(mdp,
a,
num_epochs,
epoch_length,
test_epoch_length,
max_steps,
run_tests,
value_logging=True)
print 'running experiment...'
e.run()
ak = file_utils.load_key('../access_key.key')
sk = file_utils.load_key('../secret_key.key')
bucket = 'hierarchical9'
try:
aws_util = aws_s3_utility.S3Utility(ak, sk, bucket)
aws_util.upload_directory(e.agent.logger.log_dir)
except Exception as e:
print 'error uploading to s3: {}'.format(e)
def run(learning_rate, freeze_interval, num_hidden, reg):
room_size = 5
num_rooms = 2
mdp = mdps.MazeMDP(room_size, num_rooms)
mdp.compute_states()
mdp.EXIT_REWARD = 1
mdp.MOVE_REWARD = -0.01
discount = 1
num_actions = len(mdp.get_actions(None))
batch_size = 100
print 'building network...'
network = qnetwork.QNetwork(input_shape=2 * room_size +
num_rooms**2,
batch_size=batch_size,
num_hidden_layers=2,
num_actions=4,
num_hidden=num_hidden,
discount=discount,
learning_rate=learning_rate,
regularization=reg,
update_rule='adam',
freeze_interval=freeze_interval,
rng=None)
num_epochs = 50
epoch_length = 2
test_epoch_length = 0
max_steps = 4 * (room_size * num_rooms)**2
epsilon_decay = (num_epochs * epoch_length * max_steps) / 1.5
print 'building policy...'
p = policy.EpsilonGreedy(num_actions, 0.5, 0.05, epsilon_decay)
print 'building memory...'
rm = replay_memory.ReplayMemory(batch_size, capacity=50000)
print 'building logger...'
log = logger.NeuralLogger(agent_name='QNetwork')
print 'building state adapter...'
adapter = state_adapters.CoordinatesToRowColRoomAdapter(
room_size=room_size, num_rooms=num_rooms)
# adapter = state_adapters.CoordinatesToRowColAdapter(room_size=room_size, num_rooms=num_rooms)
# adapter = state_adapters.CoordinatesToFlattenedGridAdapter(room_size=room_size, num_rooms=num_rooms)
# adapter = state_adapters.IdentityAdapter(room_size=room_size, num_rooms=num_rooms)
# adapter = state_adapters.CoordinatesToSingleRoomRowColAdapter(room_size=room_size)
print 'building agent...'
a = agent.NeuralAgent(network=network,
policy=p,
replay_memory=rm,
log=log,
state_adapter=adapter)
run_tests = False
e = experiment.Experiment(mdp,
a,
num_epochs,
epoch_length,
test_epoch_length,
max_steps,
run_tests,
value_logging=True)
e.run()
ak = file_utils.load_key('../access_key.key')
sk = file_utils.load_key('../secret_key.key')
bucket = 'hierarchical'
try:
aws_util = aws_s3_utility.S3Utility(ak, sk, bucket)
aws_util.upload_directory(e.agent.logger.log_dir)
except Exception as e:
print 'error uploading to s3: {}'.format(e)