def test_filter(env_response):
crop_low = np.array([0, 5, 10])
crop_high = np.array([5, 10, 20])
crop_filter = InputFilter()
crop_filter.add_observation_filter('observation', 'crop', ObservationCropFilter(crop_low, crop_high))
result = crop_filter.filter(env_response)[0]
unfiltered_observation = env_response.next_state['observation']
filtered_observation = result.next_state['observation']
# validate the shape of the filtered observation
assert filtered_observation.shape == (5, 5, 10)
# validate the content of the filtered observation
assert np.all(filtered_observation == unfiltered_observation[0:5, 5:10, 10:20])
# crop with -1 on some axes
crop_low = np.array([0, 0, 0])
crop_high = np.array([5, -1, -1])
crop_filter = InputFilter()
crop_filter.add_observation_filter('observation', 'crop', ObservationCropFilter(crop_low, crop_high))
result = crop_filter.filter(env_response)[0]
unfiltered_observation = env_response.next_state['observation']
filtered_observation = result.next_state['observation']
# validate the shape of the filtered observation
assert filtered_observation.shape == (5, 20, 30)
# validate the content of the filtered observation
assert np.all(filtered_observation == unfiltered_observation[0:5, :, :])
def test_filter():
# make an RGB observation smaller
env_response = EnvResponse(
next_state={'observation': np.ones([20, 30, 3])},
reward=0,
game_over=False)
rescale_filter = InputFilter()
rescale_filter.add_observation_filter(
'observation', 'rescale', ObservationRescaleSizeByFactorFilter(0.5))
result = rescale_filter.filter(env_response)[0]
unfiltered_observation = env_response.next_state['observation']
filtered_observation = result.next_state['observation']
# make sure the original observation is unchanged
assert unfiltered_observation.shape == (20, 30, 3)
# validate the shape of the filtered observation
assert filtered_observation.shape == (10, 15, 3)
# make a grayscale observation bigger
env_response = EnvResponse(next_state={'observation': np.ones([20, 30])},
reward=0,
game_over=False)
rescale_filter = InputFilter()
rescale_filter.add_observation_filter(
'observation', 'rescale', ObservationRescaleSizeByFactorFilter(2))
result = rescale_filter.filter(env_response)[0]
filtered_observation = result.next_state['observation']
# validate the shape of the filtered observation
assert filtered_observation.shape == (40, 60)
assert np.all(filtered_observation == np.ones([40, 60]))
def test_filter():
# make an RGB observation smaller
transition = EnvResponse(next_state={'observation': np.ones([20, 30, 3])},
reward=0,
game_over=False)
rescale_filter = InputFilter()
rescale_filter.add_observation_filter(
'observation', 'rescale',
ObservationRescaleToSizeFilter(
ImageObservationSpace(np.array([10, 20, 3]), high=255)))
result = rescale_filter.filter(transition)[0]
unfiltered_observation = transition.next_state['observation']
filtered_observation = result.next_state['observation']
# make sure the original observation is unchanged
assert unfiltered_observation.shape == (20, 30, 3)
# validate the shape of the filtered observation
assert filtered_observation.shape == (10, 20, 3)
assert np.all(filtered_observation == np.ones([10, 20, 3]))
# make a grayscale observation bigger
transition = EnvResponse(next_state={'observation': np.ones([20, 30])},
reward=0,
game_over=False)
rescale_filter = InputFilter()
rescale_filter.add_observation_filter(
'observation', 'rescale',
ObservationRescaleToSizeFilter(
ImageObservationSpace(np.array([40, 60]), high=255)))
result = rescale_filter.filter(transition)[0]
filtered_observation = result.next_state['observation']
# validate the shape of the filtered observation
assert filtered_observation.shape == (40, 60)
assert np.all(filtered_observation == np.ones([40, 60]))
# rescale channels -> error
# with pytest.raises(ValueError):
# InputFilter(
# observation_filters=OrderedDict([('rescale',
# ObservationRescaleToSizeFilter(ImageObservationSpace(np.array([10, 20, 1]),
# high=255)
# ))]))
# TODO: validate input to filter
# different number of axes -> error
# env_response = EnvResponse(state={'observation': np.ones([20, 30, 3])}, reward=0, game_over=False)
# rescale_filter = ObservationRescaleToSizeFilter(ObservationSpace(np.array([10, 20]))
# )
# with pytest.raises(ValueError):
# result = rescale_filter.filter(transition)
# channels first -> error
with pytest.raises(ValueError):
ObservationRescaleToSizeFilter(
ImageObservationSpace(np.array([3, 10, 20]), high=255))
def load_csv(self, csv_dataset: CsvDataset, input_filter: InputFilter) -> None:
"""
Restore the replay buffer contents from a csv file.
The csv file is assumed to include a list of transitions.
:param csv_dataset: A construct which holds the dataset parameters
:param input_filter: A filter used to filter the CSV data before feeding it to the memory.
"""
self.assert_not_frozen()
df = pd.read_csv(csv_dataset.filepath)
if len(df) > self.max_size[1]:
screen.warning("Warning! The number of transitions to load into the replay buffer ({}) is "
"bigger than the max size of the replay buffer ({}). The excessive transitions will "
"not be stored.".format(len(df), self.max_size[1]))
episode_ids = df['episode_id'].unique()
progress_bar = ProgressBar(len(episode_ids))
state_columns = [col for col in df.columns if col.startswith('state_feature')]
for e_id in episode_ids:
progress_bar.update(e_id)
df_episode_transitions = df[df['episode_id'] == e_id]
input_filter.reset()
if len(df_episode_transitions) < 2:
# we have to have at least 2 rows in each episode for creating a transition
continue
episode = Episode()
transitions = []
for (_, current_transition), (_, next_transition) in zip(df_episode_transitions[:-1].iterrows(),
df_episode_transitions[1:].iterrows()):
state = np.array([current_transition[col] for col in state_columns])
next_state = np.array([next_transition[col] for col in state_columns])
transitions.append(
Transition(state={'observation': state},
action=int(current_transition['action']), reward=current_transition['reward'],
next_state={'observation': next_state}, game_over=False,
info={'all_action_probabilities':
ast.literal_eval(current_transition['all_action_probabilities'])}),
)
transitions = input_filter.filter(transitions, deep_copy=False)
for t in transitions:
episode.insert(t)
# Set the last transition to end the episode
if csv_dataset.is_episodic:
episode.get_last_transition().game_over = True
self.store_episode(episode)
# close the progress bar
progress_bar.update(len(episode_ids))
progress_bar.close()
def test_filter_stacking():
# test that filter stacking works fine by taking as input a transition with:
# - an observation of shape 210x160,
# - a reward of 100
# filtering it by:
# - rescaling the observation to 110x84
# - cropping the observation to 84x84
# - clipping the reward to 1
# - stacking 4 observations to get 84x84x4
env_response = EnvResponse({'observation': np.ones([210, 160])}, reward=100, game_over=False)
filter1 = ObservationRescaleToSizeFilter(
output_observation_space=ImageObservationSpace(np.array([110, 84]), high=255),
)
filter2 = ObservationCropFilter(
crop_low=np.array([16, 0]),
crop_high=np.array([100, 84])
)
filter3 = RewardClippingFilter(
clipping_low=-1,
clipping_high=1
)
output_filter = ObservationStackingFilter(
stack_size=4,
stacking_axis=-1
)
input_filter = InputFilter(
observation_filters={
"observation": OrderedDict([
("filter1", filter1),
("filter2", filter2),
("output_filter", output_filter)
])},
reward_filters=OrderedDict([
("filter3", filter3)
])
)
result = input_filter.filter(env_response)[0]
observation = np.array(result.next_state['observation'])
assert observation.shape == (84, 84, 4)
assert np.all(observation == np.ones([84, 84, 4]))
assert result.reward == 1
def test_get_filtered_observation_space():
# error on wrong number of channels
rescale_filter = InputFilter()
rescale_filter.add_observation_filter(
'observation', 'rescale',
ObservationRescaleSizeByFactorFilter(
0.5, RescaleInterpolationType.BILINEAR))
observation_space = ObservationSpace(np.array([10, 20, 5]))
with pytest.raises(ValueError):
filtered_observation_space = rescale_filter.get_filtered_observation_space(
'observation', observation_space)
# error on wrong number of dimensions
observation_space = ObservationSpace(np.array([10, 20, 10, 3]))
with pytest.raises(ValueError):
filtered_observation_space = rescale_filter.get_filtered_observation_space(
'observation', observation_space)
# make sure the new observation space shape is calculated correctly
observation_space = ObservationSpace(np.array([10, 20, 3]))
filtered_observation_space = rescale_filter.get_filtered_observation_space(
'observation', observation_space)
assert np.all(filtered_observation_space.shape == np.array([5, 10, 3]))
# make sure the original observation space is unchanged
assert np.all(observation_space.shape == np.array([10, 20, 3]))
def test_filter():
rescale_filter = InputFilter(reward_filters=OrderedDict([('rescale', RewardRescaleFilter(1/10.))]))
env_response = EnvResponse(next_state={'observation': np.zeros(10)}, reward=100, game_over=False)
print(rescale_filter.observation_filters)
result = rescale_filter.filter(env_response)[0]
unfiltered_reward = env_response.reward
filtered_reward = result.reward
# validate that the reward was clipped correctly
assert filtered_reward == 10
# make sure the original reward is unchanged
assert unfiltered_reward == 100
# negative reward
env_response = EnvResponse(next_state={'observation': np.zeros(10)}, reward=-50, game_over=False)
result = rescale_filter.filter(env_response)[0]
assert result.reward == -5
def test_get_filtered_reward_space():
rescale_filter = InputFilter(reward_filters=OrderedDict([('rescale', RewardRescaleFilter(1/10.))]))
# reward is clipped
reward_space = RewardSpace(1, -100, 100)
filtered_reward_space = rescale_filter.get_filtered_reward_space(reward_space)
# make sure the new reward space shape is calculated correctly
assert filtered_reward_space.shape == 1
assert filtered_reward_space.low == -10
assert filtered_reward_space.high == 10
# unbounded rewards
reward_space = RewardSpace(1, -np.inf, np.inf)
filtered_reward_space = rescale_filter.get_filtered_reward_space(reward_space)
# make sure the new reward space shape is calculated correctly
assert filtered_reward_space.shape == 1
assert filtered_reward_space.low == -np.inf
assert filtered_reward_space.high == np.inf
def test_filter():
# make an RGB observation smaller
uint8_filter = InputFilter()
uint8_filter.add_observation_filter(
'observation', 'to_uint8',
ObservationToUInt8Filter(input_low=0, input_high=255))
observation = np.random.rand(20, 30, 3) * 255.0
env_response = EnvResponse(next_state={'observation': observation},
reward=0,
game_over=False)
result = uint8_filter.filter(env_response)[0]
unfiltered_observation = env_response.next_state['observation']
filtered_observation = result.next_state['observation']
# make sure the original observation is unchanged
assert unfiltered_observation.dtype == 'float64'
# make sure the filtering is done correctly
assert filtered_observation.dtype == 'uint8'
assert np.all(filtered_observation == observation.astype('uint8'))
def test_get_filtered_observation_space():
# Keep
observation_space = VectorObservationSpace(
3, measurements_names=['a', 'b', 'c'])
env_response = EnvResponse(next_state={'observation': np.ones([3])},
reward=0,
game_over=False)
reduction_filter = InputFilter()
reduction_filter.add_observation_filter(
'observation', 'reduce',
ObservationReductionBySubPartsNameFilter(
["a"],
ObservationReductionBySubPartsNameFilter.ReductionMethod.Keep))
filtered_observation_space = reduction_filter.get_filtered_observation_space(
'observation', observation_space)
assert np.all(filtered_observation_space.shape == np.array([1]))
assert filtered_observation_space.measurements_names == ['a']
# Discard
observation_space = VectorObservationSpace(
3, measurements_names=['a', 'b', 'c'])
env_response = EnvResponse(next_state={'observation': np.ones([3])},
reward=0,
game_over=False)
reduction_filter = InputFilter()
reduction_filter.add_observation_filter(
'observation', 'reduce',
ObservationReductionBySubPartsNameFilter(
["a"],
ObservationReductionBySubPartsNameFilter.ReductionMethod.Discard))
filtered_observation_space = reduction_filter.get_filtered_observation_space(
'observation', observation_space)
assert np.all(filtered_observation_space.shape == np.array([2]))
assert filtered_observation_space.measurements_names == ['b', 'c']
def test_get_filtered_observation_space():
# error on observation space with values not matching the filter configuration
uint8_filter = InputFilter()
uint8_filter.add_observation_filter(
'observation', 'to_uint8',
ObservationToUInt8Filter(input_low=0, input_high=200))
observation_space = ObservationSpace(np.array([1, 2, 3]), 0, 100)
with pytest.raises(ValueError):
uint8_filter.get_filtered_observation_space('observation',
observation_space)
# verify output observation space is correct
observation_space = ObservationSpace(np.array([1, 2, 3]), 0, 200)
result = uint8_filter.get_filtered_observation_space(
'observation', observation_space)
assert np.all(result.high == 255)
assert np.all(result.low == 0)
assert np.all(result.shape == observation_space.shape)
e: "{}".format(lower_under_to_upper(e) + '-v2')
for e in gym_mujoco_envs
}
mujoco_v2['walker2d'] = 'Walker2d-v2'
# Fetch
gym_fetch_envs = ['reach', 'slide', 'push', 'pick_and_place']
fetch_v1 = {
e: "{}".format('Fetch' + lower_under_to_upper(e) + '-v1')
for e in gym_fetch_envs
}
"""
Atari Environment Components
"""
AtariInputFilter = InputFilter(is_a_reference_filter=True)
AtariInputFilter.add_reward_filter('clipping', RewardClippingFilter(-1.0, 1.0))
AtariInputFilter.add_observation_filter(
'observation',
'rescaling',
ObservationRescaleToSizeFilter(
ImageObservationSpace(
np.array([84, 84, 3]), #np.array([224, 224, 3]),
high=255)))
#AtariInputFilter.add_observation_filter('observation', 'to_grayscale', ObservationRGBToYFilter())
#AtariInputFilter.add_observation_filter('observation', 'to_uint8', ObservationToUInt8Filter(0, 255))
#AtariInputFilter.add_observation_filter('observation', 'stacking', ObservationStackingFilter(4))
AtariOutputFilter = NoOutputFilter()
class Atari(GymEnvironmentParameters):
def get_graph_manager(**hp_dict):
####################
# All Default Parameters #
####################
params = {}
params["batch_size"] = int(hp_dict.get("batch_size", 64))
params["num_epochs"] = int(hp_dict.get("num_epochs", 10))
params["stack_size"] = int(hp_dict.get("stack_size", 1))
params["lr"] = float(hp_dict.get("lr", 0.0003))
params["exploration_type"] = (hp_dict.get("exploration_type",
"huber")).lower()
params["e_greedy_value"] = float(hp_dict.get("e_greedy_value", .05))
params["epsilon_steps"] = int(hp_dict.get("epsilon_steps", 10000))
params["beta_entropy"] = float(hp_dict.get("beta_entropy", .01))
params["discount_factor"] = float(hp_dict.get("discount_factor", .999))
params["loss_type"] = hp_dict.get("loss_type",
"Mean squared error").lower()
params["num_episodes_between_training"] = int(
hp_dict.get("num_episodes_between_training", 20))
params["term_cond_max_episodes"] = int(
hp_dict.get("term_cond_max_episodes", 100000))
params["term_cond_avg_score"] = float(
hp_dict.get("term_cond_avg_score", 100000))
params_json = json.dumps(params, indent=2, sort_keys=True)
print("Using the following hyper-parameters", params_json, sep='\n')
####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = TrainingSteps(
params["term_cond_max_episodes"])
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(40)
schedule_params.evaluation_steps = EnvironmentEpisodes(5)
schedule_params.heatup_steps = EnvironmentSteps(0)
#########
# Agent #
#########
agent_params = ClippedPPOAgentParameters()
agent_params.network_wrappers['main'].learning_rate = params["lr"]
agent_params.network_wrappers['main'].input_embedders_parameters[
'observation'].activation_function = 'relu'
agent_params.network_wrappers[
'main'].middleware_parameters.activation_function = 'relu'
agent_params.network_wrappers['main'].batch_size = params["batch_size"]
agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5
agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999
if params["loss_type"] == "huber":
agent_params.network_wrappers[
'main'].replace_mse_with_huber_loss = True
agent_params.algorithm.clip_likelihood_ratio_using_epsilon = 0.2
agent_params.algorithm.clipping_decay_schedule = LinearSchedule(
1.0, 0, 1000000)
agent_params.algorithm.beta_entropy = params["beta_entropy"]
agent_params.algorithm.gae_lambda = 0.95
agent_params.algorithm.discount = params["discount_factor"]
agent_params.algorithm.optimization_epochs = params["num_epochs"]
agent_params.algorithm.estimate_state_value_using_gae = True
agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentEpisodes(
params["num_episodes_between_training"])
agent_params.algorithm.num_consecutive_playing_steps = EnvironmentEpisodes(
params["num_episodes_between_training"])
agent_params.algorithm.distributed_coach_synchronization_type = DistributedCoachSynchronizationType.SYNC
if params["exploration_type"] == "categorical":
agent_params.exploration = CategoricalParameters()
else:
agent_params.exploration = EGreedyParameters()
agent_params.exploration.epsilon_schedule = LinearSchedule(
1.0, params["e_greedy_value"], params["epsilon_steps"])
###############
# Environment #
###############
SilverstoneInputFilter = InputFilter(is_a_reference_filter=True)
SilverstoneInputFilter.add_observation_filter('observation',
'to_grayscale',
ObservationRGBToYFilter())
SilverstoneInputFilter.add_observation_filter(
'observation', 'to_uint8', ObservationToUInt8Filter(0, 255))
SilverstoneInputFilter.add_observation_filter(
'observation', 'stacking',
ObservationStackingFilter(params["stack_size"]))
env_params = GymVectorEnvironment()
env_params.default_input_filter = SilverstoneInputFilter
env_params.level = 'SilverstoneRacetrack-Discrete-v0'
vis_params = VisualizationParameters()
vis_params.dump_mp4 = False
########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.test = True
preset_validation_params.min_reward_threshold = 400
preset_validation_params.max_episodes_to_achieve_reward = 1000
graph_manager = BasicRLGraphManager(
agent_params=agent_params,
env_params=env_params,
schedule_params=schedule_params,
vis_params=vis_params,
preset_validation_params=preset_validation_params)
return graph_manager, params_json
'SELECT_WEAPON3': ord("3"),
'SELECT_WEAPON4': ord("4"),
'SELECT_WEAPON5': ord("5"),
'SELECT_WEAPON6': ord("6"),
'SELECT_WEAPON7': ord("7"),
'SELECT_WEAPON8': ord("8"),
'SELECT_WEAPON9': ord("9"),
'SPEED': 304, # shift
'STRAFE': 9, # tab
'TURN180': ord("u"),
'TURN_LEFT': ord("a"), # left arrow
'TURN_RIGHT': ord("d"), # right arrow
'USE': ord("f"),
}
DoomInputFilter = InputFilter(is_a_reference_filter=True)
DoomInputFilter.add_observation_filter(
'observation', 'rescaling',
ObservationRescaleToSizeFilter(
ImageObservationSpace(np.array([60, 76, 3]), high=255)))
DoomInputFilter.add_observation_filter('observation', 'to_grayscale',
ObservationRGBToYFilter())
DoomInputFilter.add_observation_filter('observation', 'to_uint8',
ObservationToUInt8Filter(0, 255))
DoomInputFilter.add_observation_filter('observation', 'stacking',
ObservationStackingFilter(3))
DoomOutputFilter = OutputFilter(is_a_reference_filter=True)
DoomOutputFilter.add_action_filter('to_discrete', FullDiscreteActionSpaceMap())
# Starcraft Constants
_NOOP = actions.FUNCTIONS.no_op.id
_MOVE_SCREEN = actions.FUNCTIONS.Move_screen.id
_SELECT_ARMY = actions.FUNCTIONS.select_army.id
_PLAYER_RELATIVE = features.SCREEN_FEATURES.player_relative.index
_NOT_QUEUED = [0]
_SELECT_ALL = [0]
class StarcraftObservationType(Enum):
Features = 0
RGB = 1
StarcraftInputFilter = InputFilter(is_a_reference_filter=True)
StarcraftInputFilter.add_observation_filter('screen', 'move_axis',
ObservationMoveAxisFilter(0, -1))
StarcraftInputFilter.add_observation_filter(
'screen', 'rescaling',
ObservationRescaleToSizeFilter(
PlanarMapsObservationSpace(np.array([84, 84, 1]),
low=0,
high=255,
channels_axis=-1)))
StarcraftInputFilter.add_observation_filter('screen', 'to_uint8',
ObservationToUInt8Filter(0, 255))
StarcraftInputFilter.add_observation_filter('minimap', 'move_axis',
ObservationMoveAxisFilter(0, -1))
StarcraftInputFilter.add_observation_filter(
def get_graph_manager(**hp_dict):
####################
# All Default Parameters #
####################
params = {}
params["batch_size"] = int(hp_dict.get("batch_size", 64))
params["num_epochs"] = int(hp_dict.get("num_epochs", 10))
params["stack_size"] = int(hp_dict.get("stack_size", 1))
params["lr"] = float(hp_dict.get("lr", 0.0003))
params["exploration_type"] = (hp_dict.get("exploration_type", "huber")).lower()
params["e_greedy_value"] = float(hp_dict.get("e_greedy_value", .05))
params["epsilon_steps"] = int(hp_dict.get("epsilon_steps", 10000))
params["beta_entropy"] = float(hp_dict.get("beta_entropy", .01))
params["discount_factor"] = float(hp_dict.get("discount_factor", .999))
params["loss_type"] = hp_dict.get("loss_type", "Mean squared error").lower()
params["num_episodes_between_training"] = int(hp_dict.get("num_episodes_between_training", 20))
params["term_cond_max_episodes"] = int(hp_dict.get("term_cond_max_episodes", 100000))
params["term_cond_avg_score"] = float(hp_dict.get("term_cond_avg_score", 100000))
params_json = json.dumps(params, indent=2, sort_keys=True)
print("Using the following hyper-parameters", params_json, sep='\n')
####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = TrainingSteps(params["term_cond_max_episodes"])
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(40)
schedule_params.evaluation_steps = EnvironmentEpisodes(5)
schedule_params.heatup_steps = EnvironmentSteps(0)
#########
# Agent #
#########
agent_params = ClippedPPOAgentParameters()
agent_params.network_wrappers['main'].learning_rate = params["lr"]
agent_params.network_wrappers['main'].input_embedders_parameters['observation'].activation_function = 'relu'
agent_params.network_wrappers['main'].middleware_parameters.activation_function = 'relu'
agent_params.network_wrappers['main'].batch_size = params["batch_size"]
agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5
agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999
if params["loss_type"] == "huber":
agent_params.network_wrappers['main'].replace_mse_with_huber_loss = True
agent_params.algorithm.clip_likelihood_ratio_using_epsilon = 0.2
agent_params.algorithm.clipping_decay_schedule = LinearSchedule(1.0, 0, 1000000)
agent_params.algorithm.beta_entropy = params["beta_entropy"]
agent_params.algorithm.gae_lambda = 0.95
agent_params.algorithm.discount = params["discount_factor"]
agent_params.algorithm.optimization_epochs = params["num_epochs"]
agent_params.algorithm.estimate_state_value_using_gae = True
agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentEpisodes(
params["num_episodes_between_training"])
agent_params.algorithm.num_consecutive_playing_steps = EnvironmentEpisodes(params["num_episodes_between_training"])
agent_params.algorithm.distributed_coach_synchronization_type = DistributedCoachSynchronizationType.SYNC
if params["exploration_type"] == "categorical":
agent_params.exploration = CategoricalParameters()
else:
agent_params.exploration = EGreedyParameters()
agent_params.exploration.epsilon_schedule = LinearSchedule(1.0,
params["e_greedy_value"],
params["epsilon_steps"])
###############
# Environment #
###############
SilverstoneInputFilter = InputFilter(is_a_reference_filter=True)
SilverstoneInputFilter.add_observation_filter('observation', 'to_grayscale', ObservationRGBToYFilter())
SilverstoneInputFilter.add_observation_filter('observation', 'to_uint8', ObservationToUInt8Filter(0, 255))
SilverstoneInputFilter.add_observation_filter('observation', 'stacking',
ObservationStackingFilter(params["stack_size"]))
env_params = GymVectorEnvironment()
env_params.default_input_filter = SilverstoneInputFilter
env_params.level = 'SilverstoneRacetrack-Discrete-v0'
vis_params = VisualizationParameters()
vis_params.dump_mp4 = False
########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.test = True
preset_validation_params.min_reward_threshold = 400
preset_validation_params.max_episodes_to_achieve_reward = 1000
graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params,
schedule_params=schedule_params, vis_params=vis_params,
preset_validation_params=preset_validation_params)
return graph_manager, params_json
def stack_filter():
stack_filter = InputFilter()
stack_filter.add_observation_filter(
'observation', 'stack', ObservationStackingFilter(4, stacking_axis=-1))
return stack_filter
agent_params.network_wrappers['actor'].input_embedders_parameters.pop('observation')
agent_params.network_wrappers['critic'].input_embedders_parameters['measurements'] = \
agent_params.network_wrappers['critic'].input_embedders_parameters.pop('observation')
agent_params.network_wrappers['actor'].input_embedders_parameters[
'measurements'].scheme = [Dense(300)]
agent_params.network_wrappers['actor'].middleware_parameters.scheme = [
Dense(200)
]
agent_params.network_wrappers['critic'].input_embedders_parameters[
'measurements'].scheme = [Dense(400)]
agent_params.network_wrappers['critic'].middleware_parameters.scheme = [
Dense(300)
]
agent_params.network_wrappers['critic'].input_embedders_parameters[
'action'].scheme = EmbedderScheme.Empty
agent_params.input_filter = InputFilter()
agent_params.input_filter.add_reward_filter("rescale",
RewardRescaleFilter(1 / 10.))
###############
# Environment #
###############
env_params = ControlSuiteEnvironmentParameters(
level=SingleLevelSelection(control_suite_envs))
########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.trace_test_levels = ['cartpole:swingup', 'hopper:hop']
1.0, 0, 1000000)
agent_params.algorithm.beta_entropy = 0.01 # also try 0.001
agent_params.algorithm.gae_lambda = 0.95
agent_params.algorithm.discount = 0.999
agent_params.algorithm.optimization_epochs = 10
agent_params.algorithm.estimate_state_value_using_gae = True
agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentEpisodes(
20)
agent_params.algorithm.num_consecutive_playing_steps = EnvironmentEpisodes(20)
agent_params.exploration = CategoricalParameters()
agent_params.memory.max_size = (MemoryGranularity.Transitions, 10**5)
###############
# Environment #
###############
MeiroRunnerFilter = InputFilter(is_a_reference_filter=True)
env_params = GymVectorEnvironment()
env_params.level = 'RoboMaker-MeiroRunner-v0'
vis_params = VisualizationParameters()
vis_params.dump_mp4 = False
########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.test = True
preset_validation_params.min_reward_threshold = 400
preset_validation_params.max_episodes_to_achieve_reward = 1000
agent_params.algorithm.optimization_epochs = 5
agent_params.algorithm.estimate_state_value_using_gae = True
agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentEpisodes(
20)
agent_params.algorithm.num_consecutive_playing_steps = EnvironmentEpisodes(20)
#huber loss
agent_params.network_wrappers['main'].replace_mse_with_huber_loss = True
agent_params.exploration = CategoricalParameters()
agent_params.algorithm.distributed_coach_synchronization_type = DistributedCoachSynchronizationType.SYNC
###############
# Environment #
###############
SilverstoneInputFilter = InputFilter(is_a_reference_filter=True)
# SilverstoneInputFilter.add_observation_filter('left_camera', 'to_grayscale', ObservationRGBToYFilter())
# SilverstoneInputFilter.add_observation_filter('left_camera', 'to_uint8', ObservationToUInt8Filter(0, 255))
# SilverstoneInputFilter.add_observation_filter('left_camera', 'stacking', ObservationStackingFilter(1))
SilverstoneInputFilter.add_observation_filter('STEREO_CAMERAS', 'to_uint8',
ObservationToUInt8Filter(0, 255))
SilverstoneInputFilter.add_observation_filter(
'LIDAR', 'clipping', ObservationClippingFilter(0.1, 0.5))
env_params = GymVectorEnvironment()
env_params.default_input_filter = SilverstoneInputFilter
env_params.level = 'DeepRacerRacetrackCustomActionSpaceEnv-v0'
vis_params = VisualizationParameters()
vis_params.dump_mp4 = False
def rgb_to_y_filter():
rgb_to_y_filter = InputFilter()
rgb_to_y_filter.add_observation_filter('observation', 'rgb_to_y',
ObservationRGBToYFilter())
return rgb_to_y_filter
agent_params.algorithm.clip_likelihood_ratio_using_epsilon = 0.2
agent_params.algorithm.clipping_decay_schedule = LinearSchedule(1.0, 0, 1000000)
agent_params.algorithm.beta_entropy = 0.01 # also try 0.001
agent_params.algorithm.gae_lambda = 0.95
agent_params.algorithm.discount = 0.999
agent_params.algorithm.optimization_epochs = 10
agent_params.algorithm.estimate_state_value_using_gae = True
agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentEpisodes(20)
agent_params.algorithm.num_consecutive_playing_steps = EnvironmentEpisodes(20)
agent_params.exploration = CategoricalParameters()
###############
# Environment #
###############
turtlebot3_input_filter = InputFilter(is_a_reference_filter=True)
turtlebot3_input_filter.add_observation_filter('observation', 'to_grayscale', ObservationRGBToYFilter())
turtlebot3_input_filter.add_observation_filter('observation', 'to_uint8', ObservationToUInt8Filter(0, 255))
turtlebot3_input_filter.add_observation_filter('observation', 'stacking', ObservationStackingFilter(1))
env_params = GymVectorEnvironment()
env_params.default_input_filter = turtlebot3_input_filter
env_params.level = 'SageMaker-TurtleBot3-Discrete-v0'
vis_params = VisualizationParameters()
vis_params.dump_mp4 = False
########
# Test #
########
preset_validation_params = PresetValidationParameters()
TOWN1 = {"map_name": "Town01", "map_path": "/Game/Maps/Town01"}
TOWN2 = {"map_name": "Town02", "map_path": "/Game/Maps/Town02"}
key_map = {
'BRAKE': (274, ), # down arrow
'GAS': (273, ), # up arrow
'TURN_LEFT': (276, ), # left arrow
'TURN_RIGHT': (275, ), # right arrow
'GAS_AND_TURN_LEFT': (273, 276),
'GAS_AND_TURN_RIGHT': (273, 275),
'BRAKE_AND_TURN_LEFT': (274, 276),
'BRAKE_AND_TURN_RIGHT': (274, 275),
}
CarlaInputFilter = InputFilter(is_a_reference_filter=True)
CarlaInputFilter.add_observation_filter(
'forward_camera', 'rescaling',
ObservationRescaleToSizeFilter(
ImageObservationSpace(np.array([128, 180, 3]), high=255)))
CarlaInputFilter.add_observation_filter('forward_camera', 'to_grayscale',
ObservationRGBToYFilter())
CarlaInputFilter.add_observation_filter('forward_camera', 'to_uint8',
ObservationToUInt8Filter(0, 255))
CarlaInputFilter.add_observation_filter('forward_camera', 'stacking',
ObservationStackingFilter(4))
CarlaOutputFilter = NoOutputFilter()
class CameraTypes(Enum):
def get_graph_manager(**hp_dict):
####################
# All Default Parameters #
####################
params = {}
params["batch_size"] = int(hp_dict.get("batch_size", 64))
params["num_epochs"] = int(hp_dict.get("num_epochs", 10))
params["stack_size"] = int(hp_dict.get("stack_size", 1))
params["lr"] = float(hp_dict.get("lr", 0.0003))
params["lr_decay_rate"] = float(hp_dict.get("lr_decay_rate", 0))
params["lr_decay_steps"] = float(hp_dict.get("lr_decay_steps", 0))
params["exploration_type"] = (hp_dict.get("exploration_type", "categorical")).lower()
params["e_greedy_value"] = float(hp_dict.get("e_greedy_value", .05))
params["epsilon_steps"] = int(hp_dict.get("epsilon_steps", 10000))
params["beta_entropy"] = float(hp_dict.get("beta_entropy", .01))
params["discount_factor"] = float(hp_dict.get("discount_factor", .999))
params["loss_type"] = hp_dict.get("loss_type", "Mean squared error").lower()
params["num_episodes_between_training"] = int(hp_dict.get("num_episodes_between_training", 20))
params["term_cond_max_episodes"] = int(hp_dict.get("term_cond_max_episodes", 100000))
params["term_cond_avg_score"] = float(hp_dict.get("term_cond_avg_score", 100000))
params["tensorboard"] = hp_dict.get("tensorboard", False)
params["dump_mp4"] = hp_dict.get("dump_mp4", False)
params["dump_gifs"] = hp_dict.get("dump_gifs", False)
params_json = json.dumps(params, indent=2, sort_keys=True)
print("Using the following hyper-parameters", params_json, sep='\n')
####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = TrainingSteps(params["term_cond_max_episodes"])
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(40)
schedule_params.evaluation_steps = EnvironmentEpisodes(5)
schedule_params.heatup_steps = EnvironmentSteps(0)
#########
# Agent #
#########
agent_params = ClippedPPOAgentParameters()
agent_params.network_wrappers['main'].learning_rate = params["lr"]
agent_params.network_wrappers['main'].learning_rate_decay_rate = params["lr_decay_rate"]
agent_params.network_wrappers['main'].learning_rate_decay_steps = params["lr_decay_steps"]
agent_params.network_wrappers['main'].input_embedders_parameters['observation'].activation_function = 'relu'
# Replace the default CNN with single layer Conv2d(32, 3, 1)
# agent_params.network_wrappers['main'].input_embedders_parameters['observation'].scheme = EmbedderScheme.Shallow
# agent_params.network_wrappers['main'].input_embedders_parameters['observation'].dropout_rate = 0.3
agent_params.network_wrappers['main'].middleware_parameters.activation_function = 'relu'
# agent_params.network_wrappers['main'].middleware_parameters.scheme = MiddlewareScheme.Shallow
# agent_params.network_wrappers['main'].middleware_parameters.dropout_rate = 0.3
agent_params.network_wrappers['main'].batch_size = params["batch_size"]
agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5
agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999
# agent_params.network_wrappers['main'].l2_regularization = 2e-5
if params["loss_type"] == "huber":
agent_params.network_wrappers['main'].replace_mse_with_huber_loss = True
agent_params.algorithm.clip_likelihood_ratio_using_epsilon = 0.2
agent_params.algorithm.clipping_decay_schedule = LinearSchedule(1.0, 0, 1000000)
agent_params.algorithm.beta_entropy = params["beta_entropy"]
agent_params.algorithm.gae_lambda = 0.95
agent_params.algorithm.discount = params["discount_factor"]
agent_params.algorithm.optimization_epochs = params["num_epochs"]
agent_params.algorithm.estimate_state_value_using_gae = True
agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentEpisodes(
params["num_episodes_between_training"])
agent_params.algorithm.num_consecutive_playing_steps = EnvironmentEpisodes(params["num_episodes_between_training"])
agent_params.algorithm.distributed_coach_synchronization_type = DistributedCoachSynchronizationType.SYNC
if params["exploration_type"] == "categorical":
agent_params.exploration = CategoricalParameters()
else:
agent_params.exploration = EGreedyParameters()
agent_params.exploration.epsilon_schedule = LinearSchedule(1.0,
params["e_greedy_value"],
params["epsilon_steps"])
###############
# Environment #
###############
DeepRacerInputFilter = InputFilter(is_a_reference_filter=True)
# Add an observation image pertubation for many aspects
# DeepRacerInputFilter.add_observation_filter('observation', 'perturb_color', ObservationColorPerturbation(0.2))
# Rescale to much smaller input when using shallow networks to avoid OOM
# DeepRacerInputFilter.add_observation_filter('observation', 'rescaling',
# ObservationRescaleToSizeFilter(ImageObservationSpace(np.array([84, 84, 3]),
# high=255)))
DeepRacerInputFilter.add_observation_filter('observation', 'to_grayscale', ObservationRGBToYFilter())
DeepRacerInputFilter.add_observation_filter('observation', 'to_uint8', ObservationToUInt8Filter(0, 255))
DeepRacerInputFilter.add_observation_filter('observation', 'stacking',
ObservationStackingFilter(params["stack_size"]))
env_params = GymVectorEnvironment()
env_params.default_input_filter = DeepRacerInputFilter
env_params.level = 'DeepRacerRacetrackCustomActionSpaceEnv-v0'
vis_params = VisualizationParameters()
vis_params.tensorboard = params["tensorboard"]
vis_params.dump_mp4 = params["dump_mp4"]
vis_params.dump_gifs = params["dump_gifs"]
# AlwaysDumpFilter, MaxDumpFilter, EveryNEpisodesDumpFilter, SelectedPhaseOnlyDumpFilter
vis_params.video_dump_filters = [AlwaysDumpFilter()]
########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.test = True
preset_validation_params.min_reward_threshold = 400
preset_validation_params.max_episodes_to_achieve_reward = 10000
graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params,
schedule_params=schedule_params, vis_params=vis_params,
preset_validation_params=preset_validation_params)
return graph_manager, params_json
BulletOutputFilter = NoOutputFilter()
class Bullet(GymEnvironmentParameters):
def __init__(self):
super().__init__()
self.frame_skip = 1
self.default_input_filter = BulletInputFilter
self.default_output_filter = BulletOutputFilter
"""
Atari Environment Components
"""
AtariInputFilter = InputFilter(is_a_reference_filter=True)
AtariInputFilter.add_reward_filter('clipping', RewardClippingFilter(-1.0, 1.0))
AtariInputFilter.add_observation_filter(
'observation', 'rescaling',
ObservationRescaleToSizeFilter(
ImageObservationSpace(np.array([84, 84, 3]), high=255)))
AtariInputFilter.add_observation_filter('observation', 'to_grayscale',
ObservationRGBToYFilter())
AtariInputFilter.add_observation_filter('observation', 'to_uint8',
ObservationToUInt8Filter(0, 255))
AtariInputFilter.add_observation_filter('observation', 'stacking',
ObservationStackingFilter(4))
AtariOutputFilter = NoOutputFilter()
class Atari(GymEnvironmentParameters):
agent_params.network_wrappers['main'].clip_gradients = 1000
agent_params.network_wrappers[
'main'].gradients_clipping_method = GradientClippingMethod.ClipByValue
###############
# Environment #
###############
import jsbsim
import gym_jsbsim
from rl_coach.filters.filter import NoInputFilter, NoOutputFilter
from rl_coach.filters.filter import InputFilter
from rl_coach.filters.observation.observation_stacking_filter import ObservationStackingFilter
vis_params = VisualizationParameters(native_rendering=True)
input = InputFilter(is_a_reference_filter=True)
input.add_observation_filter('observation', 'stacking',
ObservationStackingFilter(10))
class MyGymVectorEnvironment(GymVectorEnvironment):
def __init__(self, level=None):
super().__init__(level=level)
self.frame_skip = 1
self.default_input_filter = NoInputFilter(
) # hrmm.. my custom input filter errored out
self.default_output_filter = NoOutputFilter()
env_params = MyGymVectorEnvironment(
level='JSBSim-HeadingControlTask-Cessna172P-Shaping.EXTRA-FG-v0')
def get_graph_manager(hp_dict, agent_list, run_phase_subject):
####################
# All Default Parameters #
####################
params = {}
params["batch_size"] = int(hp_dict.get("batch_size", 64))
params["num_epochs"] = int(hp_dict.get("num_epochs", 10))
params["stack_size"] = int(hp_dict.get("stack_size", 1))
params["lr"] = float(hp_dict.get("lr", 0.0003))
params["exploration_type"] = (hp_dict.get("exploration_type",
"categorical")).lower()
params["e_greedy_value"] = float(hp_dict.get("e_greedy_value", .05))
params["epsilon_steps"] = int(hp_dict.get("epsilon_steps", 10000))
params["beta_entropy"] = float(hp_dict.get("beta_entropy", .01))
params["discount_factor"] = float(hp_dict.get("discount_factor", .999))
params["loss_type"] = hp_dict.get("loss_type",
"Mean squared error").lower()
params["num_episodes_between_training"] = int(
hp_dict.get("num_episodes_between_training", 20))
params["term_cond_max_episodes"] = int(
hp_dict.get("term_cond_max_episodes", 100000))
params["term_cond_avg_score"] = float(
hp_dict.get("term_cond_avg_score", 100000))
params_json = json.dumps(params, indent=2, sort_keys=True)
print("Using the following hyper-parameters", params_json, sep='\n')
####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = TrainingSteps(
params["term_cond_max_episodes"])
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(40)
schedule_params.evaluation_steps = EnvironmentEpisodes(5)
schedule_params.heatup_steps = EnvironmentSteps(0)
#########
# Agent #
#########
trainable_agents_list = list()
non_trainable_agents_list = list()
for agent in agent_list:
agent_params = DeepRacerAgentParams()
if agent.network_settings:
agent_params.env_agent = agent
agent_params.network_wrappers['main'].learning_rate = params["lr"]
agent_params.network_wrappers['main'].input_embedders_parameters = \
create_input_embedder(agent.network_settings['input_embedders'],
agent.network_settings['embedder_type'],
agent.network_settings['activation_function'])
agent_params.network_wrappers['main'].middleware_parameters = \
create_middle_embedder(agent.network_settings['middleware_embedders'],
agent.network_settings['embedder_type'],
agent.network_settings['activation_function'])
input_filter = InputFilter(is_a_reference_filter=True)
for observation in agent.network_settings['input_embedders'].keys(
):
if observation == Input.LEFT_CAMERA.value or observation == Input.CAMERA.value or\
observation == Input.OBSERVATION.value:
input_filter.add_observation_filter(
observation, 'to_grayscale', ObservationRGBToYFilter())
input_filter.add_observation_filter(
observation, 'to_uint8',
ObservationToUInt8Filter(0, 255))
input_filter.add_observation_filter(
observation, 'stacking', ObservationStackingFilter(1))
if observation == Input.STEREO.value:
input_filter.add_observation_filter(
observation, 'to_uint8',
ObservationToUInt8Filter(0, 255))
if observation == Input.LIDAR.value:
input_filter.add_observation_filter(
observation, 'clipping',
ObservationClippingFilter(0.15, 1.0))
if observation == Input.SECTOR_LIDAR.value:
input_filter.add_observation_filter(
observation, 'binary', ObservationBinarySectorFilter())
agent_params.input_filter = input_filter()
agent_params.network_wrappers['main'].batch_size = params[
"batch_size"]
agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5
agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999
if params["loss_type"] == "huber":
agent_params.network_wrappers[
'main'].replace_mse_with_huber_loss = True
agent_params.algorithm.clip_likelihood_ratio_using_epsilon = 0.2
agent_params.algorithm.clipping_decay_schedule = LinearSchedule(
1.0, 0, 1000000)
agent_params.algorithm.beta_entropy = params["beta_entropy"]
agent_params.algorithm.gae_lambda = 0.95
agent_params.algorithm.discount = params["discount_factor"]
agent_params.algorithm.optimization_epochs = params["num_epochs"]
agent_params.algorithm.estimate_state_value_using_gae = True
agent_params.algorithm.num_steps_between_copying_online_weights_to_target = \
EnvironmentEpisodes(params["num_episodes_between_training"])
agent_params.algorithm.num_consecutive_playing_steps = \
EnvironmentEpisodes(params["num_episodes_between_training"])
agent_params.algorithm.distributed_coach_synchronization_type = \
DistributedCoachSynchronizationType.SYNC
if params["exploration_type"] == "categorical":
agent_params.exploration = CategoricalParameters()
else:
agent_params.exploration = EGreedyParameters()
agent_params.exploration.epsilon_schedule = LinearSchedule(
1.0, params["e_greedy_value"], params["epsilon_steps"])
trainable_agents_list.append(agent_params)
else:
non_trainable_agents_list.append(agent)
###############
# Environment #
###############
env_params = DeepRacerRacetrackEnvParameters()
env_params.agents_params = trainable_agents_list
env_params.non_trainable_agents = non_trainable_agents_list
env_params.level = 'DeepRacerRacetrackEnv-v0'
env_params.run_phase_subject = run_phase_subject
vis_params = VisualizationParameters()
vis_params.dump_mp4 = False
########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.test = True
preset_validation_params.min_reward_threshold = 400
preset_validation_params.max_episodes_to_achieve_reward = 10000
graph_manager = MultiAgentGraphManager(
agents_params=trainable_agents_list,
env_params=env_params,
schedule_params=schedule_params,
vis_params=vis_params,
preset_validation_params=preset_validation_params)
return graph_manager, params_json
def test_filter():
# make an RGB observation smaller
squeeze_filter = InputFilter()
squeeze_filter.add_observation_filter('observation', 'squeeze',
ObservationSqueezeFilter())
squeeze_filter_with_axis = InputFilter()
squeeze_filter_with_axis.add_observation_filter(
'observation', 'squeeze', ObservationSqueezeFilter(2))
observation = np.random.rand(20, 30, 1, 3)
env_response = EnvResponse(next_state={'observation': observation},
reward=0,
game_over=False)
result = squeeze_filter.filter(env_response)[0]
result_with_axis = squeeze_filter_with_axis.filter(env_response)[0]
unfiltered_observation_shape = env_response.next_state['observation'].shape
filtered_observation_shape = result.next_state['observation'].shape
filtered_observation_with_axis_shape = result_with_axis.next_state[
'observation'].shape
# make sure the original observation is unchanged
assert unfiltered_observation_shape == observation.shape
# make sure the filtering is done correctly
assert filtered_observation_shape == (20, 30, 3)
assert filtered_observation_with_axis_shape == (20, 30, 3)
observation = np.random.rand(1, 30, 1, 3)
env_response = EnvResponse(next_state={'observation': observation},
reward=0,
game_over=False)
result = squeeze_filter.filter(env_response)[0]
assert result.next_state['observation'].shape == (30, 3)
agent_params.memory.goals_space = GoalsSpace(goal_name='achieved_goal',
reward_type=ReachingGoal(distance_from_goal_threshold=0.05,
goal_reaching_reward=0,
default_reward=-1),
distance_metric=GoalsSpace.DistanceMetric.Euclidean)
agent_params.memory.shared_memory = True
# exploration parameters
agent_params.exploration = EGreedyParameters()
agent_params.exploration.epsilon_schedule = ConstantSchedule(0.3)
agent_params.exploration.evaluation_epsilon = 0
# they actually take the noise_schedule to be 0.2 * max_abs_range which is 0.1 * total_range
agent_params.exploration.continuous_exploration_policy_parameters.noise_schedule = ConstantSchedule(0.1)
agent_params.exploration.continuous_exploration_policy_parameters.evaluation_noise = 0
agent_params.input_filter = InputFilter()
agent_params.input_filter.add_observation_filter('observation', 'clipping', ObservationClippingFilter(-200, 200))
agent_params.pre_network_filter = InputFilter()
agent_params.pre_network_filter.add_observation_filter('observation', 'normalize_observation',
ObservationNormalizationFilter(name='normalize_observation'))
agent_params.pre_network_filter.add_observation_filter('achieved_goal', 'normalize_achieved_goal',
ObservationNormalizationFilter(name='normalize_achieved_goal'))
agent_params.pre_network_filter.add_observation_filter('desired_goal', 'normalize_desired_goal',
ObservationNormalizationFilter(name='normalize_desired_goal'))
###############
# Environment #
###############
env_params = GymVectorEnvironment(level=SingleLevelSelection(fetch_v1))
env_params.custom_reward_threshold = -49
def test_get_filtered_observation_space():
# error on observation space with shape not matching the filter squeeze axis configuration
squeeze_filter = InputFilter()
squeeze_filter.add_observation_filter('observation', 'squeeze',
ObservationSqueezeFilter(axis=3))
observation_space = ObservationSpace(np.array([20, 1, 30, 3]), 0, 100)
small_observation_space = ObservationSpace(np.array([20, 1, 30]), 0, 100)
with pytest.raises(ValueError):
squeeze_filter.get_filtered_observation_space('observation',
observation_space)
squeeze_filter.get_filtered_observation_space('observation',
small_observation_space)
# verify output observation space is correct
observation_space = ObservationSpace(np.array([1, 2, 3, 1]), 0, 200)
result = squeeze_filter.get_filtered_observation_space(
'observation', observation_space)
assert np.all(result.shape == np.array([1, 2, 3]))
squeeze_filter = InputFilter()
squeeze_filter.add_observation_filter('observation', 'squeeze',
ObservationSqueezeFilter())
result = squeeze_filter.get_filtered_observation_space(
'observation', observation_space)
assert np.all(result.shape == np.array([2, 3]))
agent_params.algorithm.gae_lambda = 0.95
agent_params.algorithm.discount = 0.999
agent_params.algorithm.optimization_epochs = 10
agent_params.algorithm.estimate_state_value_using_gae = True
agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentEpisodes(
20)
agent_params.algorithm.num_consecutive_playing_steps = EnvironmentEpisodes(20)
agent_params.exploration = CategoricalParameters()
agent_params.algorithm.distributed_coach_synchronization_type = DistributedCoachSynchronizationType.SYNC
###############
# Environment #
###############
SilverstoneInputFilter = InputFilter(is_a_reference_filter=True)
SilverstoneInputFilter.add_observation_filter('observation', 'to_grayscale',
ObservationRGBToYFilter())
SilverstoneInputFilter.add_observation_filter('observation', 'to_uint8',
ObservationToUInt8Filter(0, 255))
SilverstoneInputFilter.add_observation_filter('observation', 'stacking',
ObservationStackingFilter(1))
env_params = GymVectorEnvironment()
env_params.default_input_filter = SilverstoneInputFilter
env_params.level = 'DeepRacerRacetrackCustomActionSpaceEnv-v0'
vis_params = VisualizationParameters()
vis_params.dump_mp4 = False
def test_filter():
# Keep
observation_space = VectorObservationSpace(
3, measurements_names=['a', 'b', 'c'])
env_response = EnvResponse(next_state={'observation': np.ones([3])},
reward=0,
game_over=False)
reduction_filter = InputFilter()
reduction_filter.add_observation_filter(
'observation', 'reduce',
ObservationReductionBySubPartsNameFilter(
["a"],
ObservationReductionBySubPartsNameFilter.ReductionMethod.Keep))
reduction_filter.get_filtered_observation_space('observation',
observation_space)
result = reduction_filter.filter(env_response)[0]
unfiltered_observation = env_response.next_state['observation']
filtered_observation = result.next_state['observation']
# make sure the original observation is unchanged
assert unfiltered_observation.shape == (3, )
# validate the shape of the filtered observation
assert filtered_observation.shape == (1, )
# Discard
reduction_filter = InputFilter()
reduction_filter.add_observation_filter(
'observation', 'reduce',
ObservationReductionBySubPartsNameFilter(
["a"],
ObservationReductionBySubPartsNameFilter.ReductionMethod.Discard))
reduction_filter.get_filtered_observation_space('observation',
observation_space)
result = reduction_filter.filter(env_response)[0]
unfiltered_observation = env_response.next_state['observation']
filtered_observation = result.next_state['observation']
# make sure the original observation is unchanged
assert unfiltered_observation.shape == (3, )
# validate the shape of the filtered observation
assert filtered_observation.shape == (2, )