def get_config(mode='train'):
"""Generate environment config.
Args:
mode: 'train' or 'test'.
Returns:
config: Dictionary defining task/environment configuration. Can be fed as
kwargs to environment.Environment.
"""
# Select clusters to use, and their c0 factor distribution.
c0_clusters = [CLUSTERS_DISTS[cluster] for cluster in MODES[mode]]
print('Clustering task: {}, #sprites: {}'.format(MODES[mode],
NUM_SPRITES_PER_CLUSTER))
other_factors = distribs.Product([
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9),
distribs.Discrete('shape', ['square', 'triangle', 'circle']),
distribs.Discrete('scale', [0.13]),
distribs.Continuous('c1', 0.3, 1.),
distribs.Continuous('c2', 0.9, 1.),
])
# Generate the sprites to be used in this task, by combining Hue with the
# other factors.
sprite_factors = [
distribs.Product((other_factors, c0)) for c0 in c0_clusters
]
# Convert to sprites, generating the appropriate number per cluster.
sprite_gen_per_cluster = [
sprite_generators.generate_sprites(factors,
num_sprites=NUM_SPRITES_PER_CLUSTER)
for factors in sprite_factors
]
# Concat clusters into single scene to generate.
sprite_gen = sprite_generators.chain_generators(*sprite_gen_per_cluster)
# Randomize sprite ordering to eliminate any task information from occlusions
sprite_gen = sprite_generators.shuffle(sprite_gen)
# Clustering task will define rewards
task = tasks.Clustering(c0_clusters,
terminate_bonus=0.,
reward_range=10.,
sparse_reward=True)
config = {
'task': task,
'action_space': common.noisy_action_space(),
'renderers': common.renderers(),
'init_sprites': sprite_gen,
'max_episode_length': MAX_EPISODE_LENGTH,
'metadata': {
'name': os.path.basename(__file__),
'mode': mode
}
}
return config
def get_config(mode='train'):
"""Generate environment config.
Args:
mode: 'train' or 'test'.
Returns:
config: Dictionary defining task/environment configuration. Can be fed as
kwargs to environment.Environment.
"""
shared_factors = distribs.Product([
distribs.Discrete('shape', ['square', 'triangle', 'circle']),
distribs.Discrete('scale', [0.13]),
distribs.Continuous('c1', 0.3, 1.),
distribs.Continuous('c2', 0.9, 1.),
])
target_hue = distribs.Continuous('c0', 0., 0.4)
distractor_hue = distribs.Continuous('c0', 0.5, 0.9)
target_factors = distribs.Product([
MODES_TARGET_POSITIONS[mode],
target_hue,
shared_factors,
])
distractor_factors = distribs.Product([
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9),
distractor_hue,
shared_factors,
])
target_sprite_gen = sprite_generators.generate_sprites(
target_factors, num_sprites=NUM_TARGETS)
distractor_sprite_gen = sprite_generators.generate_sprites(
distractor_factors, num_sprites=NUM_DISTRACTORS)
sprite_gen = sprite_generators.chain_generators(target_sprite_gen,
distractor_sprite_gen)
# Randomize sprite ordering to eliminate any task information from occlusions
sprite_gen = sprite_generators.shuffle(sprite_gen)
task = tasks.FindGoalPosition(filter_distrib=target_hue,
terminate_distance=TERMINATE_DISTANCE)
config = {
'task': task,
'action_space': common.action_space(),
'renderers': common.renderers(),
'init_sprites': sprite_gen,
'max_episode_length': 20,
'metadata': {
'name': os.path.basename(__file__),
'mode': mode
}
}
return config
def get_config(mode='train'):
"""Generate environment config.
Args:
mode: 'train' or 'test'.
Returns:
config: Dictionary defining task/environment configuration. Can be fed as
kwargs to environment.Environment.
"""
factors = distribs.Product([
MODES_SHAPES[mode],
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9),
distribs.Discrete('scale', [0.2]),
distribs.Discrete('c0', [0.9, 0.55, 0.27], probs=[0.333, 0.334,
0.333]),
distribs.Discrete('c1', [0.6]),
distribs.Continuous('c2', 0.9, 1.),
])
sprite_gen = sprite_generators.generate_sprites(factors,
num_sprites=NUM_TARGETS,
fix_colors=True)
# Randomize sprite ordering to eliminate any task information from occlusions
sprite_gen = sprite_generators.shuffle(sprite_gen)
task = tasks.FindGoalPosition(terminate_distance=TERMINATE_DISTANCE,
sparse_reward=True)
config = {
'task':
task,
'action_space':
common.noisy_action_space(MOTION_STD_DEV, PROPORTIONAL_MOTION_NOISE,
None),
'renderers':
common.renderers(),
'init_sprites':
sprite_gen,
'max_episode_length':
60,
'metadata': {
'name': os.path.basename(__file__),
'mode': mode
}
}
return config
def get_config(mode=None):
"""Generate environment config.
Args:
mode: Unused.
Returns:
config: Dictionary defining task/environment configuration. Can be fed as
kwargs to environment.Environment.
"""
del mode # No train/test split for pure exploration
factors = distribs.Product([
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9),
distribs.Discrete('shape', ['square', 'triangle', 'circle']),
distribs.Discrete('scale', [0.13]),
distribs.Continuous('c0', 0., 1.),
distribs.Continuous('c1', 0.3, 1.),
distribs.Continuous('c2', 0.9, 1.),
])
num_sprites = lambda: np.random.randint(1, 7)
sprite_gen = sprite_generators.generate_sprites(factors,
num_sprites=num_sprites)
task = tasks.NoReward()
config = {
'task': task,
'action_space': common.action_space(),
'renderers': common.renderers(),
'init_sprites': sprite_gen,
'max_episode_length': 10,
'metadata': {
'name': os.path.basename(__file__),
'mode': mode
}
}
return config
def get_config(mode='train'):
"""Generate environment config.
Args:
mode: 'train' or 'test'.
Returns:
config: Dictionary defining task/environment configuration. Can be fed as
kwargs to environment.Environment.
"""
# Create the subtasks and their corresponding sprite generators
subtasks = []
sprite_gen_per_subtask = []
for subtask in SUBTASKS:
subtasks.append(
tasks.FindGoalPosition(
filter_distrib=subtask['distrib'],
goal_position=subtask['goal_position'],
terminate_distance=TERMINATE_DISTANCE,
raw_reward_multiplier=RAW_REWARD_MULTIPLIER))
factors = distribs.Product((
subtask['distrib'],
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9),
distribs.Discrete('shape', ['square', 'triangle', 'circle']),
distribs.Discrete('scale', [0.13]),
distribs.Continuous('c1', 0.3, 1.),
distribs.Continuous('c2', 0.9, 1.),
))
sprite_gen_per_subtask.append(
sprite_generators.generate_sprites(factors, num_sprites=1))
# Consider all combinations of subtasks
subtask_combos = list(
itertools.combinations(np.arange(len(SUBTASKS)), NUM_TARGETS))
if mode == 'train':
# Randomly sample a combination of subtasks, holding one combination out
sprite_gen = sprite_generators.sample_generator([
sprite_generators.chain_generators(
*[sprite_gen_per_subtask[i] for i in c])
for c in subtask_combos[1:]
])
elif mode == 'test':
# Use the held-out subtask combination for testing
sprite_gen = sprite_generators.chain_generators(
*[sprite_gen_per_subtask[i] for i in subtask_combos[0]])
else:
raise ValueError('Invalide mode {}.'.format(mode))
# Randomize sprite ordering to eliminate any task information from occlusions
sprite_gen = sprite_generators.shuffle(sprite_gen)
task = tasks.MetaAggregated(subtasks,
reward_aggregator='sum',
termination_criterion='all')
config = {
'task': task,
'action_space': common.action_space(),
'renderers': common.renderers(),
'init_sprites': sprite_gen,
'max_episode_length': MAX_EPISODE_LENGTH,
'metadata': {
'name': os.path.basename(__file__),
'mode': mode
}
}
return config