def testOutput(self):
g_0 = sprite_generators.generate_sprites(_distrib_0, num_sprites=1)
g_1 = sprite_generators.generate_sprites(_distrib_1, num_sprites=1)
g_chain = sprite_generators.sample_generator((g_0, g_1))
sprite_list = g_chain()
self.assertIsInstance(sprite_list, list)
self.assertLen(sprite_list, 1)
self.assertNotEqual(_distrib_0.contains(sprite_list[0].factors),
_distrib_1.contains(sprite_list[0].factors))
def testOutput(self):
g_0 = sprite_generators.generate_sprites(_distrib_0, num_sprites=1)
g_1 = sprite_generators.generate_sprites(_distrib_1, num_sprites=2)
g_chain = sprite_generators.chain_generators(g_0, g_1)
sprite_list = g_chain()
self.assertIsInstance(sprite_list, list)
self.assertLen(sprite_list, 3)
self.assertTrue(_distrib_0.contains(sprite_list[0].factors))
self.assertTrue(_distrib_1.contains(sprite_list[1].factors))
self.assertTrue(_distrib_1.contains(sprite_list[2].factors))
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 testGenerateSpritesLengthType(self, num_sprites):
g = sprite_generators.generate_sprites(_distrib_0,
num_sprites=num_sprites)
sprite_list = g()
self.assertIsInstance(sprite_list, list)
self.assertLen(sprite_list, num_sprites)
self.assertIsInstance(sprite_list[0], sprite.Sprite)
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 testGenerateSpritesCallableNum(self):
minval = 3
maxval = 6
num_sprites = np.random.randint(minval, maxval)
g = sprite_generators.generate_sprites(_distrib_0,
num_sprites=num_sprites)
sprite_list = g()
self.assertGreaterEqual(len(sprite_list), minval)
self.assertLess(len(sprite_list), maxval)
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=None):
"""Generate environment config.
Args:
mode: Unused task mode.
Returns:
config: Dictionary defining task/environment configuration. Can be fed as
kwargs to environment.Environment.
"""
del mode
shared_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.),
])
target_hue = distribs.Continuous('c0', 0., 0.4)
distractor_hue = distribs.Continuous('c0', 0.5, 0.9)
target_factors = distribs.Product([
target_hue,
shared_factors,
])
distractor_factors = distribs.Product([
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)
# Create the agent body
agent_body_factors = distribs.Product([
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9),
distribs.Discrete('shape', ['circle']),
distribs.Discrete('scale', [0.07]),
distribs.Discrete('c0', [1.]),
distribs.Discrete('c1', [0.]),
distribs.Discrete('c2', [1.]),
])
agent_body_gen = sprite_generators.generate_sprites(agent_body_factors,
num_sprites=1)
sprite_gen = sprite_generators.chain_generators(sprite_gen, agent_body_gen)
task = tasks.FindGoalPosition(filter_distrib=target_hue,
terminate_distance=TERMINATE_DISTANCE)
renderers = {
'image':
spriteworld_renderers.PILRenderer(image_size=(64, 64),
anti_aliasing=5,
color_to_rgb=color_maps.hsv_to_rgb)
}
config = {
'task': task,
'action_space': action_spaces.Embodied(step_size=0.05),
'renderers': renderers,
'init_sprites': sprite_gen,
'max_episode_length': 50,
'metadata': {
'name': os.path.basename(__file__),
'mode': mode
}
}
return config
def get_config(mode='train'):
"""Generate environment config.
Args:
mode: Unused task mode.
Returns:
config: Dictionary defining task/environment configuration. Can be fed as
kwargs to environment.Environment.
"""
# Factor distributions common to all objects.
common_factors = distribs.Product([
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9),
distribs.Continuous('angle', 0, 360, dtype='int32'),
])
# train/test split for goal-finding object scales and clustering object colors
goal_finding_scale_test = distribs.Continuous('scale', 0.08, 0.12)
green_blue_colors = distribs.Product([
distribs.Continuous('c1', 64, 256, dtype='int32'),
distribs.Continuous('c2', 64, 256, dtype='int32'),
])
if mode == 'train':
goal_finding_scale = distribs.SetMinus(
distribs.Continuous('scale', 0.05, 0.15),
goal_finding_scale_test,
)
cluster_colors = distribs.Product([
distribs.Continuous('c0', 128, 256, dtype='int32'),
green_blue_colors
])
elif mode == 'test':
goal_finding_scale = goal_finding_scale_test
cluster_colors = distribs.Product([
distribs.Continuous('c0', 0, 128, dtype='int32'), green_blue_colors
])
else:
raise ValueError(
'Invalid mode {}. Mode must be "train" or "test".'.format(mode))
# Create clustering sprite generators
sprite_gen_list = []
cluster_shapes = [
distribs.Discrete('shape', [s])
for s in ['triangle', 'square', 'pentagon']
]
for shape in cluster_shapes:
factors = distribs.Product([
common_factors,
cluster_colors,
shape,
distribs.Continuous('scale', 0.08, 0.12),
])
sprite_gen_list.append(
sprite_generators.generate_sprites(factors, num_sprites=2))
# Create goal-finding sprite generators
goal_finding_colors = [
distribs.Product([
distribs.Continuous('c0', 192, 256, dtype='int32'),
distribs.Continuous('c1', 0, 128, dtype='int32'),
distribs.Continuous('c2', 64, 128, dtype='int32'),
]),
distribs.Product([
distribs.Continuous('c0', 0, 128, dtype='int32'),
distribs.Continuous('c1', 192, 256, dtype='int32'),
distribs.Continuous('c2', 64, 128, dtype='int32'),
])
]
# Goal positions corresponding to the colors in goal_finding_colors
goal_finding_positions = [(0., 0.5), (1., 0.5)]
goal_finding_shapes = distribs.Discrete('shape', ['spoke_4', 'star_4'])
for colors in goal_finding_colors:
factors = distribs.Product([
common_factors,
goal_finding_scale,
goal_finding_shapes,
colors,
])
sprite_gen_list.append(
sprite_generators.generate_sprites(
factors, num_sprites=lambda: np.random.randint(1, 3)))
# Create distractor sprite generator
distractor_factors = distribs.Product([
common_factors,
distribs.Discrete('shape', ['circle']),
distribs.Continuous('c0', 64, 256, dtype='uint8'),
distribs.Continuous('c1', 64, 256, dtype='uint8'),
distribs.Continuous('c2', 64, 256, dtype='uint8'),
distribs.Continuous('scale', 0.08, 0.12),
])
sprite_gen_list.append(
sprite_generators.generate_sprites(
distractor_factors, num_sprites=lambda: np.random.randint(0, 3)))
# Concat clusters into single scene to generate
sprite_gen = sprite_generators.chain_generators(*sprite_gen_list)
# Randomize sprite ordering to eliminate any task information from occlusions
sprite_gen = sprite_generators.shuffle(sprite_gen)
# Create the combined task of goal-finding and clustering
task_list = []
task_list.append(
tasks.Clustering(cluster_shapes, terminate_bonus=0., reward_range=10.))
for colors, goal_pos in zip(goal_finding_colors, goal_finding_positions):
goal_finding_task = tasks.FindGoalPosition(distribs.Product(
[colors, goal_finding_shapes]),
goal_position=goal_pos,
weights_dimensions=(1, 0),
terminate_distance=0.15,
raw_reward_multiplier=30)
task_list.append(goal_finding_task)
task = tasks.MetaAggregated(task_list,
reward_aggregator='sum',
termination_criterion='all')
renderers = {
'image':
spriteworld_renderers.PILRenderer(image_size=(64, 64), anti_aliasing=5)
}
config = {
'task': task,
'action_space': action_spaces.SelectMove(scale=0.5),
'renderers': renderers,
'init_sprites': sprite_gen,
'max_episode_length': 50,
'metadata': {
'name': os.path.basename(__file__),
'mode': mode
}
}
return config
def testOutput(self):
g = sprite_generators.generate_sprites(_distrib_0, num_sprites=5)
g_shuffle = sprite_generators.shuffle(g)
sprite_list = g_shuffle()
self.assertIsInstance(sprite_list, list)
self.assertLen(sprite_list, 5)
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
