def testTerminateBonus(self, successes, reward_aggregator, terminate_bonus,
reward):
task = tasks.MetaAggregated(self.subtasks,
reward_aggregator=reward_aggregator,
terminate_bonus=terminate_bonus)
sprites, _ = self._get_sprites_and_reward_list(successes)
self.assertAlmostEqual(task.reward(sprites), reward, delta=0.1)
def testSum(self, termination_criterion, successes, success):
task = tasks.MetaAggregated(
self.subtasks,
reward_aggregator='sum',
termination_criterion=termination_criterion)
sprites, reward_list = self._get_sprites_and_reward_list(successes)
self.assertAlmostEqual(task.reward(sprites),
sum(reward_list),
delta=0.1)
self.assertEqual(task.success(sprites), success)
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 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
def testMean(self, successes):
task = tasks.MetaAggregated(self.subtasks, reward_aggregator='mean')
sprites, reward_list = self._get_sprites_and_reward_list(successes)
self.assertAlmostEqual(task.reward(sprites),
np.mean(reward_list),
delta=0.1)