def make_object_under_test(self, renderer):
self.environment = environment.Environment(
task=tasks.NoReward(),
action_space=action_spaces.SelectMove(),
renderers={'obs': renderer},
init_sprites=lambda: [sprite.Sprite(c0=255)],
max_episode_length=7)
def make_object_under_test(self):
"""Environment creator used by test_utils.EnvironmentTestMixin."""
env = environment.Environment(
task=tasks.NoReward(),
action_space=action_spaces.SelectMove(),
renderers={},
init_sprites=lambda: [sprite.Sprite(c0=255)],
max_episode_length=7)
return env
def testMoveSprites(self):
"""Take a series of actions and repeatedly check sprite motions."""
action_space = action_spaces.SelectMove(scale=0.5)
sprites = [sprite.Sprite(x=0.55, y=0.5), sprite.Sprite(x=0.5, y=0.5)]
# Move second (top) sprite
action_space.step(
np.array([0.52, 0.52, 0.5, 0.48]), sprites, keep_in_frame=False)
self.assertTrue(np.allclose(sprites[0].position, [0.55, 0.5], atol=1e-5))
self.assertTrue(np.allclose(sprites[1].position, [0.5, 0.49], atol=1e-5))
# Move first (bottom) sprite
action_space.step(
np.array([0.58, 0.5, 0.9, 0.9]), sprites, keep_in_frame=False)
self.assertTrue(np.allclose(sprites[0].position, [0.75, 0.7], atol=1e-5))
self.assertTrue(np.allclose(sprites[1].position, [0.5, 0.49], atol=1e-5))
# Move neither sprite
action_space.step(
np.array([0.58, 0.5, 0.9, 0.9]), sprites, keep_in_frame=False)
self.assertTrue(np.allclose(sprites[0].position, [0.75, 0.7], atol=1e-5))
self.assertTrue(np.allclose(sprites[1].position, [0.5, 0.49], atol=1e-5))
# Move second (top) sprite
action_space.step(
np.array([0.5, 0.5, 0.2, 0.5]), sprites, keep_in_frame=False)
self.assertTrue(np.allclose(sprites[0].position, [0.75, 0.7], atol=1e-5))
self.assertTrue(np.allclose(sprites[1].position, [0.35, 0.49], atol=1e-5))
# Move first (bottom) sprite
action_space.step(
np.array([0.78, 0.74, 0.9, 0.9]), sprites, keep_in_frame=False)
self.assertTrue(np.allclose(sprites[0].position, [0.95, 0.9], atol=1e-5))
self.assertTrue(np.allclose(sprites[1].position, [0.35, 0.49], atol=1e-5))
# Move first (bottom) sprite
action_space.step(
np.array([0.92, 0.9, 0.9, 0.5]), sprites, keep_in_frame=True)
self.assertTrue(np.allclose(sprites[0].position, [1., 0.9], atol=1e-5))
self.assertTrue(np.allclose(sprites[1].position, [0.35, 0.49], atol=1e-5))
# Move first (bottom) sprite
action_space.step(
np.array([0.98, 0.9, 0.7, 0.9]), sprites, keep_in_frame=False)
self.assertTrue(np.allclose(sprites[0].position, [1.1, 1.1], atol=1e-5))
self.assertTrue(np.allclose(sprites[1].position, [0.35, 0.49], atol=1e-5))
def testTaskTermination(self):
task = tasks.FindGoalPosition(goal_position=(0.5, 0.5))
action_space = action_spaces.SelectMove()
env_renderers = {}
init_sprites = lambda: [sprite.Sprite(x=0.25, y=0.25, c0=255)]
env = environment.Environment(task, action_space, env_renderers,
init_sprites)
donothing_action = np.array([0.25, 0.25, 0.5, 0.5])
success_action = np.array([0.25, 0.25, 0.75, 0.75])
timestep = env.step(donothing_action)
self.assertTrue(timestep.first())
timestep = env.step(donothing_action)
self.assertTrue(timestep.mid())
timestep = env.step(success_action)
self.assertTrue(timestep.last())
timestep = env.step(success_action)
self.assertTrue(timestep.first())
def testMotionCost(self, scale, action, motion_cost, true_cost):
action_space = action_spaces.SelectMove(
scale=scale, motion_cost=motion_cost)
cost = action_space.step(action, sprites=[], keep_in_frame=False)
self.assertAlmostEqual(cost, true_cost, delta=0.01)
def testGetMotion(self, scale, action, true_motion): action_space = action_spaces.SelectMove(scale=scale) motion = action_space.get_motion(action) self.assertTrue(np.allclose(motion, true_motion, atol=1e-4))
def testNoiseScale(self, action, noise_scale): action_space = action_spaces.SelectMove(scale=0.1, noise_scale=noise_scale) action_space.apply_noise_to_action(action)
def action_space():
return action_spaces.SelectMove(scale=0.25, noise_scale=0.05)
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 action_space():
return action_spaces.SelectMove(scale=0.25)
def noisy_action_space(_noise_std, _prop_motion_noise, filter_distribs):
return action_spaces.SelectMove(
scale=0.5,
noise_scale=_noise_std,
proportional_motion_noise=_prop_motion_noise,
filter_distribs=filter_distribs)
