def __init__(self,
level,
config,
num_action_repeats,
seed,
runfiles_path=None,
level_cache=None):
self._num_action_repeats = num_action_repeats
self._random_state = np.random.RandomState(seed=seed)
if runfiles_path:
deepmind_lab.set_runfiles_path(runfiles_path)
config = {k: str(v) for k, v in config.items()}
self._observation_spec = ['RGB_INTERLEAVED', 'INSTR']
renderer = config['renderer']
self.benchmark_mode = config['benchmark_mode']
if self.benchmark_mode:
print(
'BENCHMARK MODE IS ON! USE THIS ONLY FOR TESTING AND THROUGHPUT MEASUREMENT!'
)
self._env = deepmind_lab.Lab(
level=level,
observations=self._observation_spec,
config=config,
level_cache=level_cache,
renderer=renderer,
)
def __init__(
self, level, mode, render_size=(64, 64), action_repeat=4,
action_set=ACTION_SET_DEFAULT, level_cache=None, seed=None,
runfiles_path=None):
assert mode in ('train', 'test')
import deepmind_lab
if runfiles_path:
print('Setting DMLab runfiles path:', runfiles_path)
deepmind_lab.set_runfiles_path(runfiles_path)
self._config = {}
self._config['width'] = render_size[0]
self._config['height'] = render_size[1]
self._config['logLevel'] = 'WARN'
if mode == 'test':
self._config['allowHoldOutLevels'] = 'true'
self._config['mixerSeed'] = 0x600D5EED
self._action_repeat = action_repeat
self._random = np.random.RandomState(seed)
self._env = deepmind_lab.Lab(
level='contributed/dmlab30/'+level,
observations=['RGB_INTERLEAVED'],
config={k: str(v) for k, v in self._config.items()},
level_cache=level_cache)
self._action_set = action_set
self._last_image = None
self._done = True
def __init__(self, game, num_action_repeats, seed, is_test, config,
action_set=DEFAULT_ACTION_SET, level_cache_dir=None):
if is_test:
config['allowHoldOutLevels'] = 'true'
# Mixer seed for evalution, see
# https://github.com/deepmind/lab/blob/master/docs/users/python_api.md
config['mixerSeed'] = 0x600D5EED
if game in games.ALL_GAMES:
game = 'contributed/dmlab30/' + game
config['datasetPath'] = FLAGS.dataset_path
self._num_action_repeats = num_action_repeats
self._random_state = np.random.RandomState(seed=seed)
if FLAGS.homepath:
deepmind_lab.set_runfiles_path(FLAGS.homepath)
self._env = deepmind_lab.Lab(
level=game,
observations=['RGB_INTERLEAVED'],
level_cache=LevelCache(level_cache_dir) if level_cache_dir else None,
config={k: str(v) for k, v in config.items()},
)
self._action_set = action_set
self.action_space = gym.spaces.Discrete(len(self._action_set))
self.observation_space = gym.spaces.Box(
low=0,
high=255,
shape=(config['height'], config['width'], 3),
dtype=np.uint8)
def __init__(self,
level,
config,
num_action_repeats,
seed,
runfiles_path=None,
level_cache=None):
# Define initial attributes
self._num_action_repeats = num_action_repeats
self._random_state = np.random.RandomState(seed=seed)
self._width = config.get('width', None)
self._height = config.get('height', None)
# Determine observation space
if FLAGS.depth:
self._observation_spec = ['RGBD_INTERLEAVED']
else:
self._observation_spec = ['RGB_INTERLEAVED']
if False:
self._observation_spec.extend(['DEBUG.POS.TRANS', 'DEBUG.POS.ROT'])
if False:
self._observation_spec.extend(['DEBUG.POS.TRANS', 'DEBUG.POS.ROT'])
# Configure deepmind environment
if runfiles_path:
deepmind_lab.set_runfiles_path(runfiles_path)
config = {k: str(v) for k, v in config.iteritems()}
self._env = deepmind_lab.Lab(
level=level,
observations=self._observation_spec,
config=config,
level_cache=level_cache,
)
def __init__(self,
screen_width=300,
screen_height=300,
runfiles_path='',
state_processor=None,
level_script='tests/empty_room_test',
frame_skip=1,
seed=None,
level_directory=''):
self.width = screen_width
self.height = screen_height
self.runfiles_path = runfiles_path
self.level_script = level_script
self.frame_skip = frame_skip
self.seed = seed
self.level_directory = level_directory
self.processor = state_processor
self.current_state = None
self.mode = 'RGB_INTERLEAVED'
config = {
'width': str(screen_width),
'height': str(screen_height),
'levelDirectory': str(level_directory)
}
if self.runfiles_path:
deepmind_lab.set_runfiles_path(self.runfiles_path)
self.env = deepmind_lab.Lab(level_script, [self.mode], config=config)
self.number_of_actions = len(self.env.action_spec())
def is_terminal(state):
# Currently, is terminal if and only if you recieve a reward
def actor_learner_thread(child_conn, level_script, config):
env = deepmind_lab.Lab(level_script, ['RGB_INTERLEAVED'], config=config)
t = 1
while T < T_max:
d_theta, d_theta_v = 0, 0
theta, theta_v = child_conn.recv() # Get most current version
t_start = t
state = env.observations()['RGB_INTERLEAVED']
while not is_terminal(state) or t-t_start == t_max:
reward, next_state = env_step()
t += 1
T += 1
R = 0 if is_terminal(state) else valNetwork(state) #Bootstrap?
for i in range(t-1, t_start, -1):
R = reward[i] + gamma*R #accumulate these
# Accumulate gradients
d_theta +=
d_theta_v +=
child_conn.send(d_theta, d_theta_v)
def run(width, height, level_script, frame_count):
"""Spins up an environment and runs the random agent."""
config = {'width': str(width), 'height': str(height)}
if __name__ == '__main__':
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--frame_count', type=int, default=10000,
help='Number of steps to run the agent')
parser.add_argument('--width', type=int, default=80,
help='Horizontal size of the observations')
parser.add_argument('--height', type=int, default=80,
help='Vertical size of the observations')
parser.add_argument('--runfiles_path', type=str, default=None,
help='Set the runfiles path to find DeepMind Lab data')
parser.add_argument('--level_script', type=str, default='tests/empty_room_test',
help='The environment level script to load')
args = parser.parse_args()
if args.runfiles_path:
deepmind_lab.set_runfiles_path(args.runfiles_path)
run(args.width, args.height, args.level_script, args.frame_count)
def __init__(self, level, config, num_action_repeats, seed,
runfiles_path=None, level_cache=None):
self._num_action_repeats = num_action_repeats
self._random_state = np.random.RandomState(seed=seed)
if runfiles_path:
deepmind_lab.set_runfiles_path(runfiles_path)
config = {k: str(v) for k, v in config.iteritems()}
self._observation_spec = ['RGB_INTERLEAVED', 'INSTR']
self._env = deepmind_lab.Lab(
level=level,
observations=self._observation_spec,
config=config,
level_cache=level_cache,
)
def __init__(self,
level,
config,
seed,
runfiles_path=None,
level_cache=None):
self._random_state = np.random.RandomState(seed=seed)
if runfiles_path:
deepmind_lab.set_runfiles_path(runfiles_path)
config = {k: str(v) for k, v in config.items()}
self._observation_spec = ['RGBD']
self._env = deepmind_lab.Lab(
level=level,
observations=self._observation_spec,
config=config,
level_cache=level_cache,
)
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--runfiles_path',
type=str,
default=None,
help='Set the runfiles path to find DeepMind Lab data')
args = parser.parse_args()
if args.runfiles_path:
deepmind_lab.set_runfiles_path(args.runfiles_path)
# Test for 1 minute of simulation time at fixed in-game frame rate.
length = 3600
fps = 60
# Benchmark at each of the following levels at the specified resolutions and
# observation specs.
for level_script in ['nav_maze_static_01', 'lt_space_bounce_hard']:
for width, height in [(84, 84), (160, 120), (320, 240)]:
for observation_spec in ['RGB', 'RGBD']:
run(length, width, height, fps, level_script, observation_spec)
layouts.add(env.observations()[MAZE_LAYOUT_OBSERVATION])
num_layouts = len(layouts)
self.assertTrue(np.isclose(num_layouts, MAZE_LAYOUT_TRIALS))
for i in six.moves.range(MAZE_LAYOUT_TRIALS):
print('phase 2: trial {} out of {}'.format(i+1, MAZE_LAYOUT_TRIALS))
env = deepmind_lab.Lab(
'tests/maze_generation_test', [MAZE_LAYOUT_OBSERVATION],
config={
'mixerSeed': '0',
})
env.reset(seed=i+1)
layouts.add(env.observations()[MAZE_LAYOUT_OBSERVATION])
self.assertEqual(len(layouts), num_layouts)
for i in six.moves.range(MAZE_LAYOUT_TRIALS):
print('phase 3: trial {} out of {}'.format(i+1, MAZE_LAYOUT_TRIALS))
env = deepmind_lab.Lab(
'tests/maze_generation_test', [MAZE_LAYOUT_OBSERVATION],
config={
'mixerSeed': '1',
})
env.reset(seed=i+1)
layouts.add(env.observations()[MAZE_LAYOUT_OBSERVATION])
self.assertTrue(np.isclose(len(layouts) - num_layouts, MAZE_LAYOUT_TRIALS))
if __name__ == '__main__':
if 'TEST_SRCDIR' in os.environ:
deepmind_lab.set_runfiles_path(
os.path.join(os.environ['TEST_SRCDIR'],
'org_deepmind_lab'))
unittest.main()
def __init__(self,
platform,
args,
action_set=DEFAULT_ACTION_SET,
main_observation='RGB_INTERLEAVED',
action_repeat=4,
noise_type='',
tv_num_images=30):
"""Creates a DMLabWrapper.
Args:
platform: Typically 'dmlab'.
args: The environment settings.
action_set: The set of discrete actions.
main_observation: The observation returned at every time step.
action_repeat: Maximum number of times to repeat an action.
This can be less at the end of an episode.
noise_type: if not empty defines what type of noise to add to the
observation. Possible values: image_action, image, noise_action, noise.
tv_num_images: number of distinct images to be used for TV purposes.
"""
homepath = args.pop('homepath')
level_name = args.pop('levelName')
observation_format = args.pop('observationFormat')
renderer = args.pop('renderer')
seed = args.pop('seed')
string_args = {key: str(value) for key, value in args.items()}
if homepath:
deepmind_lab.set_runfiles_path(os.path.join(homepath, ))
self._env = deepmind_lab.Lab(level_name, observation_format,
string_args, renderer)
self._random_state = np.random.RandomState(seed=seed)
self._env.reset(seed=self._random_state.randint(0, 2**31 - 1))
self._action_set = action_set
self._action_repeat = action_repeat
self.width = args['width']
self.height = args['height']
self._main_observation = main_observation
self._transform_observation = lambda x: x
if main_observation == 'DEBUG.CAMERA.PLAYER_VIEW_NO_RETICLE':
# This observation format is (RGB, height, width).
# Convert it to (height, width, RGB).
self._transform_observation = lambda x: np.moveaxis(x, 0, -1)
# Build a list of all the possible actions.
self._action_list = []
for action in action_set:
self._action_list.append(np.array(action, dtype=np.intc))
self._noise_type = noise_type
self._images_for_noise = []
if self._noise_type:
if 'action' in self._noise_type:
assert action_set in [
DEFAULT_ACTION_SET_WITH_IDLE, DEFAULT_ACTION_SET
]
for image in range(1, tv_num_images + 1):
image_path = '/cns/vz-d/home/raveman/images/%d.jpeg' % image
tmp_path = os.path.join(tempfile.gettempdir(),
os.path.basename(image_path))
image = cv2.imread(tmp_path, flags=cv2.IMREAD_COLOR)
image = cv2.resize(image,
(int(self.width / 2), int(self.height / 2)),
interpolation=cv2.INTER_AREA)
# imread returns BGR not RGB
image = image[Ellipsis, ::-1]
self._images_for_noise.append(image)
def __init__(self,
platform,
args,
action_set=DEFAULT_ACTION_SET,
main_observation='RGB_INTERLEAVED',
action_repeat=4,
noise_type='',
tv_num_images=30,
level_cache_dir=None,
level_cache_tar=None,
level_cache_mode=False,
optimize_env_reset=False,
debug_name=''):
"""Creates a DMLabWrapper.
Args:
platform: Typically 'dmlab'.
args: The environment settings.
action_set: The set of discrete actions.
main_observation: The observation returned at every time step.
action_repeat: Maximum number of times to repeat an action.
This can be less at the end of an episode.
noise_type: if not empty defines what type of noise to add to the
observation. Possible values: image_action, image, noise_action, noise.
tv_num_images: number of distinct images to be used for TV purposes.
"""
self._optimize_env_reset = optimize_env_reset
homepath = args.pop('homepath')
level_name = args.pop('levelName')
observation_format = args.pop('observationFormat')
renderer = args.pop('renderer')
seed = args.pop('seed')
string_args = {key: str(value) for key, value in args.items()}
if homepath:
deepmind_lab.set_runfiles_path(os.path.join(homepath, ))
deepmind_lab_kwargs = {}
if level_cache_tar is not None:
deepmind_lab_kwargs['level_cache'] = LevelCacheTar(level_cache_tar)
elif level_cache_dir is not None:
deepmind_lab_kwargs['level_cache'] = LevelCache(level_cache_dir)
self._env = deepmind_lab.Lab(level_name, observation_format,
string_args, renderer,
**deepmind_lab_kwargs)
self._random_state = np.random.RandomState(seed=seed)
self._env.reset(seed=self._random_state.randint(0, 2**31 - 1))
if level_cache_mode:
#while True:
if debug_name in ['train']:
# 20M * 1.1 / 12 / 1800 * (1 or 2)
num_resets = (1025 if optimize_env_reset else 2050)
elif debug_name in ['valid', 'test']:
# 20M * 1.1 / 12 / 256 / 25 * (2 or 3)
num_resets = (580 if optimize_env_reset else 865)
else:
assert False
print('Level caching starts ({}). Process {}'.format(
num_resets, os.getpid()))
for _ in range(num_resets):
self._env.reset(seed=self._random_state.randint(0, 2**31 - 1))
print('Level caching done ({}). Exiting process {}'.format(
num_resets, os.getpid()))
import sys
sys.exit(0)
# Enable level caching.
self._noise_random_state = np.random.RandomState(seed=seed + 345)
self._action_set = action_set
self._action_repeat = action_repeat
self.width = args['width']
self.height = args['height']
self._main_observation = main_observation
self._transform_observation = lambda x: x
if main_observation == 'DEBUG.CAMERA.PLAYER_VIEW_NO_RETICLE':
# This observation format is (RGB, height, width).
# Convert it to (height, width, RGB).
self._transform_observation = lambda x: np.moveaxis(x, 0, -1)
# Build a list of all the possible actions.
self._action_list = []
for action in action_set:
self._action_list.append(np.array(action, dtype=np.intc))
self._noise_type = noise_type
self._images_for_noise = []
if self._noise_type:
if 'action' in self._noise_type:
assert action_set in [
DEFAULT_ACTION_SET_WITH_IDLE, DEFAULT_ACTION_SET
]
if 'image' in self._noise_type:
for image in range(1, tv_num_images + 1):
#image_path = '/cns/vz-d/home/raveman/images/%d.jpeg' % image
image_path = 'tv_images/%d.jpeg' % image
#tmp_path = os.path.join(tempfile.gettempdir(),
# os.path.basename(image_path))
#image = cv2.imread(tmp_path, flags=cv2.IMREAD_COLOR)
image = cv2.imread(image_path, flags=cv2.IMREAD_COLOR)
image = cv2.resize(
image, (int(self.width / 2), int(self.height / 2)),
interpolation=cv2.INTER_AREA)
# imread returns BGR not RGB
image = image[Ellipsis, ::-1]
self._images_for_noise.append(image)
self._agent_position_history = None
self._reward_history = None
self._last_maze_layout = None
import os
import inspect
import deepmind_lab as dl
import deepmind_lab_gym as dlg
import multiprocdmlab as mpdmlab
import numpy as np
DEEPMIND_RUNFILES_PATH = os.path.dirname(inspect.getfile(dl))
DEEPMIND_SOURCE_PATH = os.path.abspath(DEEPMIND_RUNFILES_PATH + '/..' * 5)
dl.set_runfiles_path(DEEPMIND_RUNFILES_PATH)
def get_entity_layer_path(entity_layer_name):
global DEEPMIND_RUNFILES_PATH, DEEPMIND_SOURCE_PATH
mode, size, num = entity_layer_name.split('-')
path_format = '{}/assets/entityLayers/{}/{}/entityLayers/{}.entityLayer'
path = path_format.format(DEEPMIND_SOURCE_PATH, size, mode, num)
return path
def get_game_environment(mapname='training-09x09-0127',
mode='training',
multiproc=False,
random_spawn=True,
random_goal=True,
apple_prob=0.9,
episode_length=5):
mapstrings = ','.join(
open(get_entity_layer_path(m)).read() for m in mapname.split(','))
def set_runfiles_path(path):
"""Module-level function to set the path of the DeepMind Lab DSOs."""
deepmind_lab.set_runfiles_path(path)
'alpha': prior_alpha,
'beta': prior_beta,
'theta': prior_theta,
'Sigma': prior_Sigma,
'gibbs_max_iter': gibbs_max_iter,
'gibbs_loglik_eps': gibbs_loglik_eps
}
bandit = MCMCBanditSampling(A, reward_function, reward_prior,
thompsonSampling)
return bandit
if __name__ == "__main__":
path = os.path.dirname(inspect.getfile(deepmind_lab))
deepmind_lab.set_runfiles_path(path)
coords, A = DRRN.action_segments() # Rotation Axes, Number of Arms
rewards = 0
R = 100 # Number of realizations to run
t_max = 300 # Time-instants to run the bandit
width = 8
height = 8
d_context = width * height * 3 # Context dimension
K = 5 # Number of mixtures per arm of the bandit
prior_K = 5 # Assumed prior number of mixtures (per arm)
pi = np.random.rand(A, K)
pi = pi / pi.sum(axis=1, keepdims=True) # Mixture proportions per arm
theta = np.random.randn(A, K, d_context) # Thetas per arm and mixtures
sigma = np.ones((A, K)) # Variances per arm and mixtures
obs = env.observations()
action = agent.step(reward, obs[observation_spec])
reward = env.step(action, num_steps=1)
t1 = time.time()
duration = t1 - t0
print('resolution: %i x %i, spec: %s, steps: %i, duration: %.1f, fps: %.1f' %
(width, height, observation_spec, length, duration, length / duration))
if __name__ == '__main__':
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--runfiles_path', type=str, default=None,
help='Set the runfiles path to find DeepMind Lab data')
args = parser.parse_args()
if args.runfiles_path:
deepmind_lab.set_runfiles_path(args.runfiles_path)
# Test for 1 minute of simulation time at fixed in-game frame rate.
length = 3600
fps = 60
# Benchmark at each of the following levels at the specified resolutions and
# observation specs.
for level_script in ['nav_maze_static_01', 'lt_space_bounce_hard']:
for width, height in [(84, 84), (160, 120), (320, 240)]:
for observation_spec in ['RGB', 'RGBD']:
run(length, width, height, fps, level_script, observation_spec)
parser.add_argument('--level_script',
type=str,
default='tests/empty_room_test',
help='The environment level script to load')
parser.add_argument('--base_path',
type=str,
default='/om/user/prinster/lab/my_data/',
help='base_path')
parser.add_argument('--num_envs',
type=str,
default=16,
help='num environments to run in parallel')
parser.add_argument('--learning_rate',
type=str,
default=1e-3,
help='learning_rate')
parser.add_argument('--exp_name',
type=str,
default='test',
help='exp_name')
parser.add_argument(
'--slurm_array_index',
type=int,
default=0,
help='id provided by slurm for which experiment to run')
args = parser.parse_args()
if args.runfiles_path:
deepmind_lab.set_runfiles_path(args.runfiles_path)
run(args.slurm_array_index, args.base_path)
def test_discretized_random_agent_run(self, length=100):
env = deepmind_lab.Lab(
'tests/demo_map', ['RGB_INTERLACED'],
config={
'fps': '60',
'width': '80',
'height': '80'
})
env.reset()
agent = random_agent.DiscretizedRandomAgent()
reward = 0
for _ in xrange(length):
if not env.is_running():
print('Environment stopped early')
env.reset()
obs = env.observations()
action = agent.step(reward, obs['RGB_INTERLACED'])
reward = env.step(action, 1)
self.assertIsInstance(reward, float)
if __name__ == '__main__':
if os.environ.get('TEST_SRCDIR'):
deepmind_lab.set_runfiles_path(
os.path.join(os.environ['TEST_SRCDIR'],
'org_deepmind_lab'))
unittest.main()