def test_example_runner_ba_av():
sim = ExampleAVSimulator()
sim.reset(np.zeros(5))
expert_trajectory = {'state': sim.clone_state(),
'reward': 0,
'action': np.zeros(6),
'observation': np.zeros(5)}
with patch('examples.AV.example_runner_ba_av.compress_pickle.dump', side_effect=MemoryError):
with patch('examples.AV.example_runner_ba_av.LocalTFRunner.train', new=lambda x: 0):
runner(env_args={'id': 'ast_toolbox:GoExploreAST-v1'}, algo_args={'expert_trajectory': [expert_trajectory] * 50})
def __init__(self,
open_loop=True,
blackbox_sim_state=True,
fixed_init_state=False,
s_0=None,
simulator=None,
reward_function=None,
spaces=None):
# Constant hyper-params -- set by user
self.open_loop = open_loop
self.blackbox_sim_state = blackbox_sim_state # is this redundant?
self.spaces = spaces
# These are set by reset, not the user
self._done = False
self._reward = 0.0
self._info = []
self._step = 0
self._action = None
self._actions = []
self._first_step = True
self.reward_range = (-float('inf'), float('inf'))
self.metadata = None
self._cum_reward = 0.0
if s_0 is None:
self._init_state = self.observation_space.sample()
else:
self._init_state = s_0
self._fixed_init_state = fixed_init_state
self.simulator = simulator
if self.simulator is None:
self.simulator = ExampleAVSimulator()
self.reward_function = reward_function
if self.reward_function is None:
self.reward_function = ExampleAVReward()
if hasattr(self.simulator, "vec_env_executor") and callable(
getattr(self.simulator, "vec_env_executor")):
self.vectorized = True
else:
self.vectorized = False
def __init__(self,
algo,
env,
n_envs=1,
open_loop=True,
sim=ExampleAVSimulator(),
reward_function=ExampleAVReward()):
# pdb.set_trace()
self.open_loop = open_loop
self.sim = sim
self.reward_function = reward_function
super().__init__(algo, env, n_envs)
def run_task(snapshot_config, *_):
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
with tf.variable_scope('AST', reuse=tf.AUTO_REUSE):
with LocalTFRunner(
snapshot_config=snapshot_config, max_cpus=4, sess=sess) as local_runner:
# Instantiate the example classes
sim = ExampleAVSimulator(**sim_args)
reward_function = ExampleAVReward(**reward_args)
spaces = ExampleAVSpaces(**spaces_args)
# Create the environment
if 'id' in env_args:
env_args.pop('id')
env = TfEnv(normalize(ASTEnv(simulator=sim,
reward_function=reward_function,
spaces=spaces,
**env_args
)))
# Instantiate the garage objects
policy = GaussianLSTMPolicy(env_spec=env.spec, **policy_args)
baseline = LinearFeatureBaseline(env_spec=env.spec, **baseline_args)
optimizer = ConjugateGradientOptimizer
optimizer_args = {'hvp_approach': FiniteDifferenceHvp(base_eps=1e-5)}
algo = PPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
optimizer=optimizer,
optimizer_args=optimizer_args,
**algo_args)
sampler_cls = ASTVectorizedSampler
local_runner.setup(
algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args={"open_loop": False,
"sim": sim,
"reward_function": reward_function,
'n_envs': n_parallel})
# Run the experiment
local_runner.train(**runner_args)
def __init__(self, algo, env, n_envs=1, open_loop=True, batch_simulate=False,
sim=ExampleAVSimulator(), reward_function=ExampleAVReward()):
"""
"""
# pdb.set_trace()
super(BatchSampler, self).__init__(algo, env)
self.n_envs = n_envs
self.open_loop = open_loop
self.sim = sim
self.reward_function = reward_function
self.open_loop = open_loop
self.batch_simulate = batch_simulate
def run_task(snapshot_config, *_):
with LocalTFRunner(snapshot_config=snapshot_config, max_cpus=1) as runner:
# Instantiate the example classes
sim = ExampleAVSimulator()
reward_function = ExampleAVReward()
spaces = ExampleAVSpaces()
# Create the environment
env = TfEnv(
normalize(
ASTEnv(blackbox_sim_state=True,
fixed_init_state=True,
s_0=[-0.5, -4.0, 1.0, 11.17, -35.0],
simulator=sim,
reward_function=reward_function,
spaces=spaces)))
# Instantiate the garage objects
policy = GaussianLSTMPolicy(name='lstm_policy',
env_spec=env.spec,
hidden_dim=64)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TRPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=max_path_length,
discount=0.99,
kl_constraint='soft',
max_kl_step=0.01)
sampler_cls = ASTVectorizedSampler
runner.setup(algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args={
"sim": sim,
"reward_function": reward_function
})
runner.train(n_epochs=1, batch_size=4000, plot=False)
print("Installation successfully validated")
class ASTEnv(gym.Env):
# class ASTEnv(GarageEnv):
def __init__(self,
open_loop=True,
blackbox_sim_state=True,
fixed_init_state=False,
s_0=None,
simulator=None,
reward_function=None,
spaces=None):
# Constant hyper-params -- set by user
self.open_loop = open_loop
self.blackbox_sim_state = blackbox_sim_state # is this redundant?
self.spaces = spaces
# These are set by reset, not the user
self._done = False
self._reward = 0.0
self._info = []
self._step = 0
self._action = None
self._actions = []
self._first_step = True
self.reward_range = (-float('inf'), float('inf'))
self.metadata = None
self._cum_reward = 0.0
if s_0 is None:
self._init_state = self.observation_space.sample()
else:
self._init_state = s_0
self._fixed_init_state = fixed_init_state
self.simulator = simulator
if self.simulator is None:
self.simulator = ExampleAVSimulator()
self.reward_function = reward_function
if self.reward_function is None:
self.reward_function = ExampleAVReward()
if hasattr(self.simulator, "vec_env_executor") and callable(
getattr(self.simulator, "vec_env_executor")):
self.vectorized = True
else:
self.vectorized = False
# super().__init__(self)
def step(self, action):
"""
Run one timestep of the environment's dynamics. When end of episode
is reached, reset() should be called to reset the environment's internal state.
Input
-----
action : an action provided by the environment
Outputs
-------
(observation, reward, done, info)
observation : agent's observation of the current environment
reward [Float] : amount of reward due to the previous action
done : a boolean, indicating whether the episode has ended
info : a dictionary containing other diagnostic information from the previous action
"""
self._action = action
self._actions.append(action)
action_return = self._action
# Update simulation step
obs = self.simulator.step(self._action)
if (obs is None) or (self.open_loop is
True) or (self.blackbox_sim_state):
obs = np.array(self._init_state)
# if self.simulator.is_goal():
if self.simulator.is_terminal() or self.simulator.is_goal():
self._done = True
# Calculate the reward for this step
self._reward = self.reward_function.give_reward(
action=self._action, info=self.simulator.get_reward_info())
self._cum_reward += self._reward
# Update instance attributes
self._step = self._step + 1
self._simulator_state = self.simulator.clone_state()
self._env_state = np.concatenate(
(self._simulator_state, np.array([self._cum_reward
]), np.array([self._step])),
axis=0)
if self._done:
self.simulator.simulate(self._actions, self._init_state)
if not (self.simulator.is_goal() or self.simulator.is_terminal()):
pdb.set_trace()
return Step(
observation=obs,
reward=self._reward,
done=self._done,
cache=self._info,
actions=action_return,
# step = self._step -1,
# real_actions=self._action,
state=self._env_state,
# root_action=self.root_action,
is_terminal=self.simulator.is_terminal(),
is_goal=self.simulator.is_goal())
def simulate(self, actions):
if not self._fixed_init_state:
self._init_state = self.observation_space.sample()
self.simulator.simulate(actions, self._init_state)
def reset(self):
"""
Resets the state of the environment, returning an initial observation.
Outputs
-------
observation : the initial observation of the space. (Initial reward is assumed to be 0.)
"""
self._actions = []
if not self._fixed_init_state:
self._init_state = self.observation_space.sample()
self._done = False
self._reward = 0.0
self._cum_reward = 0.0
self._info = []
self._action = None
self._actions = []
self._first_step = True
self._step = 0
obs = np.array(self.simulator.reset(self._init_state))
if not self._fixed_init_state:
obs = np.concatenate((obs, np.array(self._init_state)), axis=0)
return obs
@property
def action_space(self):
"""
Returns a Space object
"""
if self.spaces is None:
# return self._to_garage_space(self.simulator.action_space)
return self.simulator.action_space
else:
return self.spaces.action_space
@property
def observation_space(self):
"""
Returns a Space object
"""
if self.spaces is None:
# return self._to_garage_space(self.simulator.observation_space)
return self.simulator.observation_space
else:
return self.spaces.observation_space
def get_cache_list(self):
return self._info
def log(self):
self.simulator.log()
def render(self):
if hasattr(self.simulator, "render") and callable(
getattr(self.simulator, "render")):
return self.simulator.render()
else:
return None
def close(self):
if hasattr(self.simulator, "close") and callable(
getattr(self.simulator, "close")):
self.simulator.close()
else:
return None
def vec_env_executor(self, n_envs, max_path_length):
return self.simulator.vec_env_executor(n_envs, max_path_length,
self.reward_function,
self._fixed_init_state,
self._init_state,
self.open_loop)
def log_diagnostics(self, paths):
pass
@cached_property
def spec(self):
"""
Returns an EnvSpec.
Returns:
spec (garage.envs.EnvSpec)
"""
return EnvSpec(observation_space=self.observation_space,
action_space=self.action_space)
def run_task(snapshot_config, *_):
config = tf.ConfigProto(device_count={'GPU': 0})
# config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
with tf.variable_scope('AST', reuse=tf.AUTO_REUSE):
# Instantiate the example classes
sim = ExampleAVSimulator(**sim_args)
# blackbox_sim_state=True,
# open_loop=False,
# fixed_initial_state=True,
# max_path_length=max_path_length)
reward_function = ExampleAVReward(**reward_args)
spaces = ExampleAVSpaces(**spaces_args)
# Create the environment
# env1 = GoExploreASTEnv(open_loop=False,
# blackbox_sim_state=True,
# fixed_init_state=True,
# s_0=[-0.5, -4.0, 1.0, 11.17, -35.0],
# simulator=sim,
# reward_function=reward_function,
# spaces=spaces
# )
# env1 = gym.make('ast_toolbox:GoExploreAST-v1',
# blackbox_sim_state=True,
# open_loop=False,
# fixed_init_state=True,
# s_0=s_0,
# simulator=sim,
# reward_function=reward_function,
# spaces=spaces
# )
env1 = gym.make(id=env_args.pop('id'),
simulator=sim,
reward_function=reward_function,
spaces=spaces,
**env_args)
env2 = normalize(env1)
env = TfEnv(env2)
# Instantiate the garage objects
policy = GoExplorePolicy(env_spec=env.spec)
baseline = LinearFeatureBaseline(env_spec=env.spec,
**baseline_args)
algo = GoExplore(env_spec=env.spec,
env=env,
policy=policy,
baseline=baseline,
**algo_args)
# db_filename=db_filename,
# max_db_size=max_db_size,
# env=env,
#
# policy=policy,
# baseline=baseline,
# # robust_policy=robust_policy,
# # robust_baseline=robust_baseline,
# max_path_length=max_path_length,
# discount=discount,
# save_paths_gap=1,
# save_paths_path=log_dir,
# # whole_paths=whole_paths
# )
sampler_cls = BatchSampler
# sampler_args = {'n_envs': n_parallel}
sampler_args = {}
with LocalTFRunner(snapshot_config=snapshot_config,
sess=sess) as local_runner:
local_runner.setup(algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args=sampler_args)
# local_runner.setup(
# algo=algo,
# env=env,
# sampler_cls=sampler_cls,
# sampler_args={"sim": sim,
# "reward_function": reward_function})
# Run the experiment
best_cell = local_runner.train(
**runner_args
) # n_epochs=n_itr, batch_size=batch_size, plot=False)
log_dir = run_experiment_args['log_dir']
db_filename = algo_args['db_filename']
s_0 = env_args['s_0']
pool_DB = db.DB()
pool_DB.open(db_filename + '_pool.dat',
dbname=None,
dbtype=db.DB_HASH,
flags=db.DB_CREATE)
d_pool = shelve.Shelf(pool_DB,
protocol=pickle.HIGHEST_PROTOCOL)
# pdb.set_trace()
print(best_cell)
temp = best_cell
paths = []
while (temp.parent is not None):
print(temp.observation)
action = temp.observation[1:].astype(np.float32) / 1000
paths.append({
'state': temp.state,
'reward': temp.reward,
'action': action,
'observation': np.array(s_0)
})
temp = d_pool[temp.parent]
print(temp.observation)
paths.append({
'state': temp.state,
'reward': temp.reward,
'action': action,
'observation': np.array(s_0)
})
# pdb.set_trace()
d_pool.close()
with open(log_dir + '/expert_trajectory.p', 'wb') as f:
pickle.dump([paths], f)
print('done!')
def run_task(snapshot_config, *_):
seed = 0
# top_k = 10
np.random.seed(seed)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
with tf.variable_scope('AST', reuse=tf.AUTO_REUSE):
with LocalTFRunner(snapshot_config=snapshot_config,
max_cpus=4,
sess=sess) as local_runner:
# Instantiate the example classes
sim = ExampleAVSimulator(**sim_args)
reward_function = ExampleAVReward(**reward_args)
spaces = ExampleAVSpaces(**spaces_args)
# Create the environment
if 'id' in env_args:
env_args.pop('id')
env = ASTEnv(simulator=sim,
reward_function=reward_function,
spaces=spaces,
**env_args)
top_paths = BPQ.BoundedPriorityQueue(**bpq_args)
if mcts_type == 'mcts':
print('mcts')
algo = MCTS(env=env, top_paths=top_paths, **algo_args)
elif mcts_type == 'mctsbv':
print('mctsbv')
algo = MCTSBV(env=env,
top_paths=top_paths,
**algo_args)
elif mcts_type == 'mctsrs':
print('mctsrs')
algo = MCTSRS(env=env,
top_paths=top_paths,
**algo_args)
else:
raise NotImplementedError
sampler_cls = ASTVectorizedSampler
local_runner.setup(algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args={
"open_loop": False,
"sim": sim,
"reward_function": reward_function,
"n_envs": n_parallel
})
# Run the experiment
local_runner.train(**runner_args)
log_dir = run_experiment_args['log_dir']
with open(log_dir + '/best_actions.p', 'rb') as f:
best_actions = pickle.load(f)
expert_trajectories = []
for actions in best_actions:
sim.reset(s_0=env_args['s_0'])
path = []
for action in actions:
obs = sim.step(action)
state = sim.clone_state()
reward = reward_function.give_reward(
action=action, info=sim.get_reward_info())
path.append({
'state': state,
'reward': reward,
'action': action,
'observation': obs
})
expert_trajectories.append(path)
with open(log_dir + '/expert_trajectory.p', 'wb') as f:
pickle.dump(expert_trajectories, f)
def run_task(snapshot_config, *_):
config = tf.ConfigProto(device_count={'GPU': 0})
# config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
with tf.variable_scope('AST', reuse=tf.AUTO_REUSE):
# Instantiate the example classes
sim = ExampleAVSimulator(**sim_args)
reward_function = ExampleAVReward(**reward_args)
spaces = ExampleAVSpaces(**spaces_args)
# Create the environment
# env1 = GoExploreASTEnv(open_loop=False,
# blackbox_sim_state=True,
# fixed_init_state=True,
# s_0=[-0.5, -4.0, 1.0, 11.17, -35.0],
# simulator=sim,
# reward_function=reward_function,
# spaces=spaces
# )
env1 = gym.make(id=env_args.pop('id'),
simulator=sim,
reward_function=reward_function,
spaces=spaces,
**env_args)
env2 = normalize(env1)
env = TfEnv(env2)
sampler_cls = BatchSampler
# sampler_args = {'n_envs': n_parallel}
sampler_args = {}
# expert_trajectory_file = log_dir + '/expert_trajectory.p'
# with open(expert_trajectory_file, 'rb') as f:
# expert_trajectory = pickle.load(f)
#
# #Run backwards algorithm to robustify
with LocalTFRunner(snapshot_config=snapshot_config,
sess=sess) as local_runner:
policy = GaussianLSTMPolicy(env_spec=env.spec,
**policy_args)
# name='lstm_policy',
# env_spec=env.spec,
# hidden_dim=64,
# use_peepholes=True)
baseline = LinearFeatureBaseline(env_spec=env.spec,
**baseline_args)
optimizer = ConjugateGradientOptimizer
optimizer_args = {
'hvp_approach': FiniteDifferenceHvp(base_eps=1e-5)
}
algo = BackwardAlgorithm(env=env,
env_spec=env.spec,
policy=policy,
baseline=baseline,
optimizer=optimizer,
optimizer_args=optimizer_args,
**algo_args)
# expert_trajectory=expert_trajectory[-1],
# epochs_per_step = 10,
# scope=None,
# max_path_length=max_path_length,
# discount=discount,
# gae_lambda=1,
# center_adv=True,
# positive_adv=False,
# fixed_horizon=False,
# pg_loss='surrogate_clip',
# lr_clip_range=1.0,
# max_kl_step=1.0,
# policy_ent_coeff=0.0,
# use_softplus_entropy=False,
# use_neg_logli_entropy=False,
# stop_entropy_gradient=False,
# entropy_method='no_entropy',
# name='PPO',
# )
local_runner.setup(algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args=sampler_args)
results = local_runner.train(**runner_args)
# pdb.set_trace()
print('done')
log_dir = run_experiment_args['log_dir']
with open(log_dir + '/paths.gz', 'wb') as f:
try:
compress_pickle.dump(results,
f,
compression="gzip",
set_default_extension=False)
except MemoryError:
print('1')
# pdb.set_trace()
for idx, result in enumerate(results):
with open(
log_dir + '/path_' + str(idx) + '.gz',
'wb') as ff:
try:
compress_pickle.dump(
result,
ff,
compression="gzip",
set_default_extension=False)
except MemoryError:
print('2')
class GoExploreASTEnv(gym.Env, Parameterized):
r"""Gym environment to turn general AST tasks into garage compatible problems with Go-Explore style resets.
Certain algorithms, such as Go-Explore and the Backwards Algorithm, require deterministic resets of the
simulator. `GoExploreASTEnv` handles this by cloning simulator states and saving them in a cell structure. The
cells are then stored in a hashed database.
Parameters
----------
open_loop : bool
True if the simulation is open-loop, meaning that AST must generate all actions ahead of time, instead
of being able to output an action in sync with the simulator, getting an observation back before
the next action is generated. False to get interactive control, which requires that `blackbox_sim_state`
is also False.
blackbox_sim_state : bool
True if the true simulation state can not be observed, in which case actions and the initial conditions are
used as the observation. False if the simulation state can be observed, in which case it will be used
fixed_init_state : bool
True if the initial state is fixed, False to sample the initial state for each rollout from the observaation
space.
s_0 : array_like
The initial state for the simulation (ignored if `fixed_init_state` is False)
simulator : :py:class:`ast_toolbox.simulators.ASTSimulator`
The simulator wrapper, inheriting from `ast_toolbox.simulators.ASTSimulator`.
reward_function : :py:class:`ast_toolbox.rewards.ASTReward`
The reward function, inheriting from `ast_toolbox.rewards.ASTReward`.
spaces : :py:class:`ast_toolbox.spaces.ASTSpaces`
The observation and action space definitions, inheriting from `ast_toolbox.spaces.ASTSpaces`.
"""
def __init__(self,
open_loop=True,
blackbox_sim_state=True,
fixed_init_state=False,
s_0=None,
simulator=None,
reward_function=None,
spaces=None):
# gym_env = gym.make('ast_toolbox:GoExploreAST-v0', {'test':'test string'})
# pdb.set_trace()
# super().__init__(gym_env)
# Constant hyper-params -- set by user
self.open_loop = open_loop
self.blackbox_sim_state = blackbox_sim_state # is this redundant?
self.spaces = spaces
if spaces is None:
self.spaces = ExampleAVSpaces()
# These are set by reset, not the user
self._done = False
self._reward = 0.0
self._info = {}
self._step = 0
self._action = None
self._actions = []
self._first_step = True
self.reward_range = (-float('inf'), float('inf'))
self.metadata = None
self.spec._entry_point = []
self._cum_reward = 0.0
self.root_action = None
self.sample_limit = 10000
self.simulator = simulator
if self.simulator is None:
self.simulator = ExampleAVSimulator()
if s_0 is None:
self._init_state = self.observation_space.sample()
else:
self._init_state = s_0
self._fixed_init_state = fixed_init_state
self.reward_function = reward_function
if self.reward_function is None:
self.reward_function = ExampleAVReward()
if hasattr(self.simulator, "vec_env_executor") and callable(getattr(self.simulator, "vec_env_executor")):
self.vectorized = True
else:
self.vectorized = False
# super().__init__(self)
# Always call Serializable constructor last
self.params_set = False
self.db_filename = 'database.dat'
self.key_list = []
self.max_value = 0
self.robustify_state = []
self.robustify = False
Parameterized.__init__(self)
def sample(self, population):
r"""Sample a cell from the cell pool with likelihood proportional to cell fitness.
The sampling is done using Stochastic Acceptance [1]_, with inspiration from John B Nelson's blog [2]_.
The sampler rejects cells until the acceptance criterea is met. If the maximum number of rejections is
exceeded, the sampler then will sample uniformly sample a cell until it finds a cell with fitness > 0. If
the second sampling phase also exceeds the rejection limit, then the function raises an exception.
Parameters
----------
population : list
A list containing the population of cells to sample from.
Returns
-------
object
The sampled cell.
Raises
------
ValueError
If the maximum number of rejections is exceeded in both the proportional and the uniform sampling phases.
References
----------
.. [1] Lipowski, Adam, and Dorota Lipowska. "Roulette-wheel selection via stochastic acceptance."
Physica A: Statistical Mechanics and its Applications 391.6 (2012): 2193-2196.
`<https://arxiv.org/pdf/1109.3627.pdf>`_
.. [2] `<https://jbn.github.io/fast_proportional_selection/>`_
"""
attempts = 0
while attempts < self.sample_limit:
attempts += 1
candidate = population[random.choice(self.p_key_list.value)]
if random.random() < (candidate.fitness / self.p_max_value.value):
return candidate
attempts = 0
while attempts < self.sample_limit:
attempts += 1
candidate = population[random.choice(self.p_key_list.value)]
if candidate.fitness > 0:
print("Returning Uniform Random Sample - Max Attempts Reached!")
return candidate
print("Failed to find a valid state for reset!")
raise ValueError
# return population[random.choice(self.p_key_list.value)]
def get_first_cell(self):
r"""Returns a the observation and state of the initial state, to be used for a root cell.
Returns
-------
obs : array_like
Agent's observation of the current environment.
state : array_like
The cloned simulation state at the current cell, used for resetting if chosen to start a rollout.
"""
obs = self.env_reset()
if self.blackbox_sim_state:
obs = self.simulator.get_first_action()
state = np.concatenate((self.simulator.clone_state(),
np.array([self._cum_reward]),
np.array([-1])),
axis=0)
return obs, state
def step(self, action):
r"""
Run one timestep of the environment's dynamics. When end of episode
is reached, reset() should be called to reset the environment's internal state.
Parameters
----------
action : array_like
An action provided by the environment.
Returns
-------
: :py:func:`garage.envs.base.Step`
A step in the rollout.
Contains the following information:
- observation (array_like): Agent's observation of the current environment.
- reward (float): Amount of reward due to the previous action.
- done (bool): Is the current step a terminal or goal state, ending the rollout.
- cache (dict): A dictionary containing other diagnostic information from the current step.
- actions (array_like): The action taken at the current.
- state (array_like): The cloned simulation state at the current cell, used for resetting if chosen to start a rollout.
- is_terminal (bool): Whether or not the current cell is a terminal state.
- is_goal (bool): Whether or not the current cell is a goal state.
"""
self._env_state_before_action = self._env_state.copy()
self._action = action
self._actions.append(action)
action_return = self._action
# Update simulation step
obs = self.simulator.step(self._action)
if (obs is None) or (self.open_loop is True) or (self.blackbox_sim_state):
obs = np.array(self._init_state)
# Add step number to differentiate identical actions
if self.simulator.is_terminal() or self.simulator.is_goal():
self._done = True
# Calculate the reward for this step
self._reward = self.reward_function.give_reward(
action=self._action,
info=self.simulator.get_reward_info())
self._cum_reward += self._reward
# Update instance attributes
self._step = self._step + 1
self._simulator_state = self.simulator.clone_state()
self._env_state = np.concatenate((self._simulator_state,
np.array([self._cum_reward]),
np.array([self._step])),
axis=0)
return Step(observation=obs,
reward=self._reward,
done=self._done,
cache=self._info,
actions=action_return,
state=self._env_state_before_action,
root_action=self.root_action,
is_terminal=self.simulator.is_terminal(),
is_goal=self.simulator.is_goal())
def simulate(self, actions):
r"""Run a full simulation rollout.
Parameters
----------
actions : list[array_like]
A list of array_likes, where each member is the action taken at that step.
Returns
-------
int
The step of the trajectory where a collision was found, or -1 if a collision was not found.
dict
A dictionary of simulation information for logging and diagnostics.
"""
if not self._fixed_init_state:
self._init_state = self.observation_space.sample()
return self.simulator.simulate(actions, self._init_state)
def reset(self, **kwargs):
r"""Resets the state of the environment, returning an initial observation.
The reset has 2 modes.
In the "robustify" mode (self.p_robustify_state.value is not None), the simulator resets
the environment to `p_robustify_state.value`. It then returns the initial condition.
In the "Go-Explore" mode, the environment attempts to sample a cell from the cell pool. If successful,
the simulator is reset to the cell's state. On an error, the environment is reset to the intial state.
Returns
-------
observation : array_like
The initial observation of the space. (Initial reward is assumed to be 0.)
"""
try:
# print(self.p_robustify_state.value)
if self.p_robustify_state is not None and self.p_robustify_state.value is not None and len(
self.p_robustify_state.value) > 0:
state = self.p_robustify_state.value
# print('-----------Robustify Init-----------------')
# print('-----------Robustify Init: ', state, ' -----------------')
self.simulator.restore_state(state[:-2])
obs = self.simulator.observation_return()
self._done = False
self._cum_reward = state[-2]
self._step = state[-1]
# pdb.set_trace()
self.robustify = True
self._simulator_state = self.simulator.clone_state()
self._env_state = np.concatenate((self._simulator_state,
np.array([self._cum_reward]),
np.array([self._step])),
axis=0)
return self._init_state
flag = db.DB_RDONLY
pool_DB = db.DB()
pool_DB.open(self.p_db_filename.value, dbname=None, dbtype=db.DB_HASH, flags=flag)
dd_pool = shelve.Shelf(pool_DB, protocol=pickle.HIGHEST_PROTOCOL)
cell = self.sample(dd_pool)
dd_pool.close()
pool_DB.close()
if cell.state is not None:
# pdb.set_trace()
if np.all(cell.state == 0):
print("-------DEFORMED CELL STATE-------")
obs = self.env_reset()
else:
self.simulator.restore_state(cell.state[:-2])
if self.simulator.is_terminal() or self.simulator.is_goal():
print('-------SAMPLED TERMINAL STATE-------')
pdb.set_trace()
obs = self.env_reset()
else:
if cell.score == 0.0 and cell.parent is not None:
print("Reset to cell with score 0.0 ---- terminal: ", self.simulator.is_terminal(),
" goal: ", self.simulator.is_goal(), " obs: ", cell.observation)
obs = self.simulator.observation_return()
self._done = False
self._cum_reward = cell.state[-2]
self._step = cell.state[-1]
self.root_action = cell.observation
else:
print("Reset from start")
obs = self.env_reset()
self._simulator_state = self.simulator.clone_state()
self._env_state = np.concatenate((self._simulator_state,
np.array([self._cum_reward]),
np.array([self._step])),
axis=0)
# pdb.set_trace()
except db.DBBusyError:
print("DBBusyError")
obs = self.env_reset()
except db.DBLockNotGrantedError or db.DBLockDeadlockError:
print("db.DBLockNotGrantedError or db.DBLockDeadlockError")
obs = self.env_reset()
except db.DBForeignConflictError:
print("DBForeignConflictError")
obs = self.env_reset()
except db.DBAccessError:
print("DBAccessError")
obs = self.env_reset()
except db.DBPermissionsError:
print("DBPermissionsError")
obs = self.env_reset()
except db.DBNoSuchFileError:
print("DBNoSuchFileError")
obs = self.env_reset()
except db.DBError:
print("DBError")
obs = self.env_reset()
except BaseException:
print("Failed to get state from database")
pdb.set_trace()
obs = self.env_reset()
return obs
def env_reset(self):
r"""Resets the state of the environment, returning an initial observation.
Returns
-------
observation : array_like
The initial observation of the space. (Initial reward is assumed to be 0.)
"""
self._actions = []
if not self._fixed_init_state:
self._init_state = self.observation_space.sample()
self._done = False
self._reward = 0.0
self._cum_reward = 0.0
self._info = {'actions': []}
self._action = self.simulator.get_first_action()
self._actions = []
self._first_step = True
self._step = 0
obs = np.array(self.simulator.reset(self._init_state))
if not self.blackbox_sim_state:
obs = np.concatenate((obs, np.array(self._init_state)), axis=0)
self.root_action = self._action
return obs
@property
def action_space(self):
r"""Convenient access to the environment's action space.
Returns
-------
: `gym.spaces.Space <https://gym.openai.com/docs/#spaces>`_
The action space of the reinforcement learning problem.
"""
if self.spaces is None:
# return self._to_garage_space(self.simulator.action_space)
return self.simulator.action_space
else:
return self.spaces.action_space
@property
def observation_space(self):
r"""Convenient access to the environment's observation space.
Returns
-------
: `gym.spaces.Space <https://gym.openai.com/docs/#spaces>`_
The observation space of the reinforcement learning problem.
"""
if self.spaces is None:
# return self._to_garage_space(self.simulator.observation_space)
return self.simulator.observation_space
else:
return self.spaces.observation_space
def get_cache_list(self):
"""Returns the environment info cache.
Returns
-------
dict
A dictionary containing diagnostic and logging information for the environment.
"""
return self._info
def log(self):
r"""Calls the simulator's `log` function.
"""
self.simulator.log()
def render(self, **kwargs):
r"""Calls the simulator's `render` function, if it exists.
Returns
-------
None or object
Returns the output of the simulator's `render` function, or None if the simulator has no `render` function.
"""
if hasattr(self.simulator, "render") and callable(getattr(self.simulator, "render")):
return self.simulator.render(**kwargs)
else:
return None
def close(self):
r"""Calls the simulator's `close` function, if it exists.
Returns
-------
None or object
Returns the output of the simulator's `close` function, or None if the simulator has no `close` function.
"""
if hasattr(self.simulator, "close") and callable(getattr(self.simulator, "close")):
self.simulator.close()
else:
return None
@cached_property
def spec(self):
r"""Returns a garage environment specification.
Returns
-------
:py:class:`garage.envs.env_spec.EnvSpec`
A garage environment specification.
"""
return EnvSpec(
observation_space=self.observation_space,
action_space=self.action_space)
def get_params_internal(self, **tags):
r"""Returns the parameters associated with the given tags.
Parameters
----------
tags : dict[bool]
For each tag, a parameter is returned if the parameter name matches the tag's key
Returns
-------
list
List of parameters
"""
# this lasagne function also returns all var below the passed layers
if not self.params_set:
self.p_db_filename = GoExploreParameter("db_filename", self.db_filename)
self.p_key_list = GoExploreParameter("key_list", self.key_list)
self.p_max_value = GoExploreParameter("max_value", self.max_value)
self.p_robustify_state = GoExploreParameter("robustify_state", self.robustify_state)
self.params_set = True
if tags.pop("db_filename", False):
return [self.p_db_filename]
if tags.pop("key_list", False):
return [self.p_key_list]
if tags.pop("max_value", False):
return [self.p_max_value]
if tags.pop("robustify_state", False):
return [self.p_robustify_state]
return [self.p_db_filename, self.p_key_list, self.p_max_value, self.p_robustify_state] # , self.p_downsampler]
def set_param_values(self, param_values, **tags):
r"""Set the values of parameters
Parameters
----------
param_values : object
Value to set the parameter to.
tags : dict[bool]
For each tag, a parameter is returned if the parameter name matches the tag's key
"""
debug = tags.pop("debug", False)
for param, value in zip(
self.get_params(**tags),
param_values):
param.set_value(value)
if debug:
print("setting value of %s" % param.name)
def get_param_values(self, **tags):
"""Return the values of internal parameters.
Parameters
----------
tags : dict[bool]
For each tag, a parameter is returned if the parameter name matches the tag's key
Returns
-------
list
A list of parameter values.
"""
return [
param.get_value(borrow=True) for param in self.get_params(**tags)
]
def downsample(self, obs):
"""Create a downsampled approximation of the observed simulation state.
Parameters
----------
obs : array_like
The observed simulation state.
Returns
-------
array_like
The downsampled approximation of the observed simulation state.
"""
return obs
def run_task(snapshot_config, *_):
seed = 0
# top_k = 10
np.random.seed(seed)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
with tf.variable_scope('AST', reuse=tf.AUTO_REUSE):
with LocalTFRunner(
snapshot_config=snapshot_config, max_cpus=4, sess=sess) as local_runner:
# Instantiate the example classes
sim = ExampleAVSimulator(**sim_args)
reward_function = ExampleAVReward(**reward_args)
spaces = ExampleAVSpaces(**spaces_args)
# Create the environment
if 'id' in env_args:
env_args.pop('id')
env = ASTEnv(simulator=sim,
reward_function=reward_function,
spaces=spaces,
**env_args
)
top_paths = BPQ.BoundedPriorityQueue(**bpq_args)
if mcts_type == 'mcts':
print('mcts')
algo = MCTS(env=env,
top_paths=top_paths,
**algo_args)
elif mcts_type == 'mctsbv':
print('mctsbv')
algo = MCTSBV(env=env,
top_paths=top_paths,
**algo_args)
elif mcts_type == 'mctsrs':
print('mctsrs')
algo = MCTSRS(env=env,
top_paths=top_paths,
**algo_args)
else:
raise NotImplementedError
sampler_cls = ASTVectorizedSampler
sampler_args['sim'] = sim
sampler_args['reward_function'] = reward_function
local_runner.setup(algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args=sampler_args)
# Run the experiment
local_runner.train(**runner_args)
log_dir = run_experiment_args['log_dir']
if save_expert_trajectory:
load_convert_and_save_mcts_expert_trajectory(
best_actions_filename=log_dir + '/best_actions.p',
expert_trajectory_filename=log_dir + '/expert_trajectory.p',
sim=sim,
s_0=env_args['s_0'],
reward_function=reward_function)
def run_task(snapshot_config, *_):
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
with tf.variable_scope('AST', reuse=tf.AUTO_REUSE):
with LocalTFRunner(snapshot_config=snapshot_config,
max_cpus=4,
sess=sess) as local_runner:
# Instantiate the example classes
sim = ExampleAVSimulator(**sim_args)
reward_function = ExampleAVReward(**reward_args)
spaces = ExampleAVSpaces(**spaces_args)
# Create the environment
if 'id' in env_args:
env_args.pop('id')
env = TfEnv(
normalize(
ASTEnv(simulator=sim,
reward_function=reward_function,
spaces=spaces,
**env_args)))
# Instantiate the garage objects
policy = GaussianLSTMPolicy(env_spec=env.spec,
**policy_args)
baseline = LinearFeatureBaseline(env_spec=env.spec,
**baseline_args)
optimizer = ConjugateGradientOptimizer
optimizer_args = {
'hvp_approach': FiniteDifferenceHvp(base_eps=1e-5)
}
algo = PPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
optimizer=optimizer,
optimizer_args=optimizer_args,
**algo_args)
sampler_cls = ASTVectorizedSampler
sampler_args['sim'] = sim
sampler_args['reward_function'] = reward_function
local_runner.setup(algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args=sampler_args)
# Run the experiment
local_runner.train(**runner_args)
if save_expert_trajectory:
load_convert_and_save_drl_expert_trajectory(
last_iter_filename=os.path.join(
run_experiment_args['log_dir'], 'itr_' +
str(runner_args['n_epochs'] - 1) + '.pkl'),
expert_trajectory_filename=os.path.join(
run_experiment_args['log_dir'],
'expert_trajectory.pkl'))
print('done!')
class ASTEnv(gym.Env):
r""" Gym environment to turn general AST tasks into garage compatible problems.
Parameters
----------
open_loop : bool
True if the simulation is open-loop, meaning that AST must generate all actions ahead of time, instead
of being able to output an action in sync with the simulator, getting an observation back before
the next action is generated. False to get interactive control, which requires that `blackbox_sim_state`
is also False.
blackbox_sim_state : bool
True if the true simulation state can not be observed, in which case actions and the initial conditions are
used as the observation. False if the simulation state can be observed, in which case it will be used.
fixed_init_state : bool
True if the initial state is fixed, False to sample the initial state for each rollout from the observaation
space.
s_0 : array_like
The initial state for the simulation (ignored if `fixed_init_state` is False)
simulator : :py:class:`ast_toolbox.simulators.ASTSimulator`
The simulator wrapper, inheriting from `ast_toolbox.simulators.ASTSimulator`.
reward_function : :py:class:`ast_toolbox.rewards.ASTReward`
The reward function, inheriting from `ast_toolbox.rewards.ASTReward`.
spaces : :py:class:`ast_toolbox.spaces.ASTSpaces`
The observation and action space definitions, inheriting from `ast_toolbox.spaces.ASTSpaces`.
"""
def __init__(self,
open_loop=True,
blackbox_sim_state=True,
fixed_init_state=False,
s_0=None,
simulator=None,
reward_function=None,
spaces=None):
# Constant hyper-params -- set by user
self.open_loop = open_loop
self.blackbox_sim_state = blackbox_sim_state # is this redundant?
self.spaces = spaces
# These are set by reset, not the user
self._done = False
self._reward = 0.0
self._step = 0
self._action = None
self._actions = []
self._first_step = True
self.reward_range = (-float('inf'), float('inf'))
self.metadata = None
self._cum_reward = 0.0
if s_0 is None:
self._init_state = self.observation_space.sample()
else:
self._init_state = s_0
self._fixed_init_state = fixed_init_state
self.simulator = simulator
if self.simulator is None:
self.simulator = ExampleAVSimulator()
self.reward_function = reward_function
if self.reward_function is None:
self.reward_function = ExampleAVReward()
if hasattr(self.simulator, "vec_env_executor") and callable(getattr(self.simulator, "vec_env_executor")):
self.vectorized = True
else:
self.vectorized = False
# super().__init__(self)
def step(self, action):
r"""
Run one timestep of the environment's dynamics. When end of episode
is reached, reset() should be called to reset the environment's internal state.
Parameters
----------
action : array_like
An action provided by the environment.
Returns
-------
: :py:func:`garage.envs.base.Step`
A step in the rollout.
Contains the following information:
- observation (array_like): Agent's observation of the current environment.
- reward (float): Amount of reward due to the previous action.
- done (bool): Is the current step a terminal or goal state, ending the rollout.
- actions (array_like): The action taken at the current.
- state (array_like): The cloned simulation state at the current cell, used for resetting if chosen to start a rollout.
- is_terminal (bool): Whether or not the current cell is a terminal state.
- is_goal (bool): Whether or not the current cell is a goal state.
"""
self._env_state_before_action = self._env_state.copy()
self._action = action
self._actions.append(action)
action_return = self._action
# Update simulation step
obs = self.simulator.step(self._action)
if (obs is None) or (self.open_loop is True) or (self.blackbox_sim_state):
obs = np.array(self._init_state)
if self.simulator.is_terminal() or self.simulator.is_goal():
self._done = True
# Calculate the reward for this step
self._reward = self.reward_function.give_reward(
action=self._action,
info=self.simulator.get_reward_info())
self._cum_reward += self._reward
# Update instance attributes
self._step = self._step + 1
self._simulator_state = self.simulator.clone_state()
self._env_state = np.concatenate((self._simulator_state,
np.array([self._cum_reward]),
np.array([self._step])),
axis=0)
return Step(observation=obs,
reward=self._reward,
done=self._done,
actions=action_return,
state=self._env_state_before_action,
is_terminal=self.simulator.is_terminal(),
is_goal=self.simulator.is_goal())
def simulate(self, actions):
r"""Run a full simulation rollout.
Parameters
----------
actions : list[array_like]
A list of array_likes, where each member is the action taken at that step.
Returns
-------
int
The step of the trajectory where a collision was found, or -1 if a collision was not found.
dict
A dictionary of simulation information for logging and diagnostics.
"""
if not self._fixed_init_state:
self._init_state = self.observation_space.sample()
return self.simulator.simulate(actions, self._init_state)
def reset(self):
r"""Resets the state of the environment, returning an initial observation.
Returns
-------
observation : array_like
The initial observation of the space. (Initial reward is assumed to be 0.)
"""
self._actions = []
if not self._fixed_init_state:
self._init_state = self.observation_space.sample()
self._done = False
self._reward = 0.0
self._cum_reward = 0.0
self._action = None
self._actions = []
self._first_step = True
self._step = 0
obs = np.array(self.simulator.reset(self._init_state))
if not self._fixed_init_state:
obs = np.concatenate((obs, np.array(self._init_state)), axis=0)
self._simulator_state = self.simulator.clone_state()
self._env_state = np.concatenate((self._simulator_state,
np.array([self._cum_reward]),
np.array([self._step])),
axis=0)
return obs
@property
def action_space(self):
r"""Convenient access to the environment's action space.
Returns
-------
: `gym.spaces.Space <https://gym.openai.com/docs/#spaces>`_
The action space of the reinforcement learning problem.
"""
if self.spaces is None:
# return self._to_garage_space(self.simulator.action_space)
return self.simulator.action_space
else:
return self.spaces.action_space
@property
def observation_space(self):
r"""Convenient access to the environment's observation space.
Returns
-------
: `gym.spaces.Space <https://gym.openai.com/docs/#spaces>`_
The observation space of the reinforcement learning problem.
"""
if self.spaces is None:
# return self._to_garage_space(self.simulator.observation_space)
return self.simulator.observation_space
else:
return self.spaces.observation_space
def log(self):
r"""Calls the simulator's `log` function.
"""
self.simulator.log()
def render(self, **kwargs):
r"""Calls the simulator's `render` function, if it exists.
Parameters
----------
kwargs :
Keyword arguments used in the simulators `render` function.
Returns
-------
None or object
Returns the output of the simulator's `render` function, or None if the simulator has no `render` function.
"""
if hasattr(self.simulator, "render") and callable(getattr(self.simulator, "render")):
return self.simulator.render(**kwargs)
else:
return None
def close(self):
r"""Calls the simulator's `close` function, if it exists.
Returns
-------
None or object
Returns the output of the simulator's `close` function, or None if the simulator has no `close` function.
"""
if hasattr(self.simulator, "close") and callable(getattr(self.simulator, "close")):
return self.simulator.close()
else:
return None
@cached_property
def spec(self):
r"""Returns a garage environment specification.
Returns
-------
:py:class:`garage.envs.env_spec.EnvSpec`
A garage environment specification.
"""
return EnvSpec(
observation_space=self.observation_space,
action_space=self.action_space)
class GoExploreASTEnv(gym.Env, Parameterized):
def __init__(self,
open_loop=True,
blackbox_sim_state=True,
fixed_init_state=False,
s_0=None,
simulator=None,
reward_function=None,
spaces=None):
# gym_env = gym.make('ast_toolbox:GoExploreAST-v0', {'test':'test string'})
# pdb.set_trace()
# super().__init__(gym_env)
# Constant hyper-params -- set by user
self.open_loop = open_loop
self.blackbox_sim_state = blackbox_sim_state # is this redundant?
self.spaces = spaces
if spaces is None:
self.spaces = ExampleAVSpaces()
# These are set by reset, not the user
self._done = False
self._reward = 0.0
self._info = {}
self._step = 0
self._action = None
self._actions = []
self._first_step = True
self.reward_range = (-float('inf'), float('inf'))
self.metadata = None
self.spec._entry_point = []
self._cum_reward = 0.0
self.root_action = None
self.sample_limit = 10000
self.simulator = simulator
if self.simulator is None:
self.simulator = ExampleAVSimulator()
if s_0 is None:
self._init_state = self.observation_space.sample()
else:
self._init_state = s_0
self._fixed_init_state = fixed_init_state
self.reward_function = reward_function
if self.reward_function is None:
self.reward_function = ExampleAVReward()
if hasattr(self.simulator, "vec_env_executor") and callable(
getattr(self.simulator, "vec_env_executor")):
self.vectorized = True
else:
self.vectorized = False
# super().__init__(self)
# Always call Serializable constructor last
self.params_set = False
self.db_filename = 'database.dat'
self.key_list = []
self.max_value = 0
self.robustify_state = []
self.robustify = False
Parameterized.__init__(self)
def sample(self, population):
# Proportional sampling: Stochastic Acceptance
# https://arxiv.org/pdf/1109.3627.pdf
# https://jbn.github.io/fast_proportional_selection/
attempts = 0
while attempts < self.sample_limit:
attempts += 1
candidate = population[random.choice(self.p_key_list.value)]
if random.random() < (candidate.fitness / self.p_max_value.value):
return candidate
attempts = 0
while attempts < self.sample_limit:
attempts += 1
candidate = population[random.choice(self.p_key_list.value)]
if candidate.fitness > 0:
print(
"Returning Uniform Random Sample - Max Attempts Reached!")
return candidate
print("Failed to find a valid state for reset!")
raise ValueError
# return population[random.choice(self.p_key_list.value)]
def get_first_cell(self):
# obs = self.env.env.reset()
# state = self.env.env.clone_state()
obs = self.env_reset()
if self.blackbox_sim_state:
# obs = self.downsample(self.simulator.get_first_action())
obs = self.simulator.get_first_action()
# else:
# obs = self.env_reset()
# obs = self.simulator.reset(self._init_state)
state = np.concatenate(
(self.simulator.clone_state(), np.array([self._cum_reward
]), np.array([-1])),
axis=0)
# pdb.set_trace()
return obs, state
def step(self, action):
"""
Run one timestep of the environment's dynamics. When end of episode
is reached, reset() should be called to reset the environment's internal state.
Input
-----
action : an action provided by the environment
Outputs
-------
(observation, reward, done, info)
observation : agent's observation of the current environment
reward [Float] : amount of reward due to the previous action
done : a boolean, indicating whether the episode has ended
info : a dictionary containing other diagnostic information from the previous action
"""
self._action = action
self._actions.append(action)
action_return = self._action
# Update simulation step
obs = self.simulator.step(self._action)
if (obs is None) or (self.open_loop is
True) or (self.blackbox_sim_state):
# print('Open Loop:', obs)
obs = np.array(self._init_state)
# if not self.robustify:
# action_return = self.downsample(action_return)
# if self.simulator.is_goal():
# Add step number to differentiate identical actions
# obs = np.concatenate((np.array([self._step]), self.downsample(obs)), axis=0)
if self.simulator.is_terminal() or self.simulator.is_goal():
self._done = True
# Calculate the reward for this step
self._reward = self.reward_function.give_reward(
action=self._action, info=self.simulator.get_reward_info())
self._cum_reward += self._reward
# Update instance attributes
self._step = self._step + 1
self._simulator_state = self.simulator.clone_state()
self._env_state = np.concatenate(
(self._simulator_state, np.array([self._cum_reward
]), np.array([self._step])),
axis=0)
# if self.robustify:
# # No obs?
# # pdb.set_trace()
# obs = self._init_state
# else:
# # print(self.robustify_state)
# obs = self.downsample(obs)
# pdb.set_trace()
return Step(
observation=obs,
reward=self._reward,
done=self._done,
cache=self._info,
actions=action_return,
# step = self._step -1,
# real_actions=self._action,
state=self._env_state,
root_action=self.root_action,
is_terminal=self.simulator.is_terminal(),
is_goal=self.simulator.is_goal())
def simulate(self, actions):
if not self._fixed_init_state:
self._init_state = self.observation_space.sample()
return self.simulator.simulate(actions, self._init_state)
def reset(self, **kwargs):
"""
This method is necessary to suppress a deprecated warning
thrown by gym.Wrapper.
Calls reset on wrapped env.
"""
try:
# print(self.p_robustify_state.value)
if self.p_robustify_state is not None and self.p_robustify_state.value is not None and len(
self.p_robustify_state.value) > 0:
state = self.p_robustify_state.value
print('-----------Robustify Init-----------------')
print('-----------Robustify Init: ', state,
' -----------------')
self.simulator.restore_state(state[:-2])
obs = self.simulator._get_obs()
self._done = False
self._cum_reward = state[-2]
self._step = state[-1]
# pdb.set_trace()
self.robustify = True
return self._init_state
# pdb.set_trace()
# start = time.time()
flag = db.DB_RDONLY
pool_DB = db.DB()
# tick1 = time.time()
pool_DB.open(self.p_db_filename.value,
dbname=None,
dbtype=db.DB_HASH,
flags=flag)
# tick2 = time.time()
dd_pool = shelve.Shelf(pool_DB, protocol=pickle.HIGHEST_PROTOCOL)
# tick3 = time.time()
# keys = dd_pool.keys()
# tick4_1 = time.time()
# list_of_keys = list(keys)
# tick4_2 = time.time()
# choice = random.choice(self.p_key_list.value)
# import pdb; pdb.set_trace()
# tick4_3 = time.time()
# cell = dd_pool[choice]
cell = self.sample(dd_pool)
# tick5 = time.time()
dd_pool.close()
# tick6 = time.time()
pool_DB.close()
# tick7 = time.time()
# print("Make DB: ", 100*(tick1 - start)/(tick7 - start), " %")
# print("Open DB: ", 100*(tick2 - tick1) / (tick7 - start), " %")
# print("Open Shelf: ", 100*(tick4_2 - tick2) / (tick7 - start), " %")
# # print("Get all keys: ", 100*(tick4_1 - tick3) / (tick7 - start), " %")
# # print("Make list of all keys: ", 100 * (tick4_2 - tick4_1) / (tick7 - start), " %")
# print("Choose random cell: ", 100 * (tick4_3 - tick4_2) / (tick7 - start), " %")
# print("Get random cell: ", 100*(tick5 - tick4_3) / (tick7 - start), " %")
# print("Close shelf: ", 100*(tick6 - tick5) / (tick7 - start), " %")
# print("Close DB: ", 100*(tick7 - tick6) / (tick7 - start), " %")
# print("DB Access took: ", time.time() - start, " s")
if cell.state is not None:
# pdb.set_trace()
if np.all(cell.state == 0):
print("-------DEFORMED CELL STATE-------")
obs = self.env_reset()
else:
# print("restore state: ", cell.state)
self.simulator.restore_state(cell.state[:-2])
if self.simulator.is_terminal() or self.simulator.is_goal(
):
print('-------SAMPLED TERMINAL STATE-------')
pdb.set_trace()
obs = self.env_reset()
else:
# print("restored")
if cell.score == 0.0 and cell.parent is not None:
print(
"Reset to cell with score 0.0 ---- terminal: ",
self.simulator.is_terminal(), " goal: ",
self.simulator.is_goal(), " obs: ",
cell.observation)
obs = self.simulator._get_obs()
self._done = False
self._cum_reward = cell.state[-2]
self._step = cell.state[-1]
self.root_action = cell.observation
# print("restore obs: ", obs)
else:
print("Reset from start")
obs = self.env_reset()
# pdb.set_trace()
except db.DBBusyError:
print("DBBusyError")
obs = self.env_reset()
except db.DBLockNotGrantedError or db.DBLockDeadlockError:
print("db.DBLockNotGrantedError or db.DBLockDeadlockError")
obs = self.env_reset()
except db.DBForeignConflictError:
print("DBForeignConflictError")
obs = self.env_reset()
except db.DBAccessError:
print("DBAccessError")
obs = self.env_reset()
except db.DBPermissionsError:
print("DBPermissionsError")
obs = self.env_reset()
except db.DBNoSuchFileError:
print("DBNoSuchFileError")
obs = self.env_reset()
except db.DBError:
print("DBError")
obs = self.env_reset()
except BaseException:
print("Failed to get state from database")
pdb.set_trace()
obs = self.env_reset()
return obs
def env_reset(self):
"""
Resets the state of the environment, returning an initial observation.
Outputs
-------
observation : the initial observation of the space. (Initial reward is assumed to be 0.)
"""
self._actions = []
if not self._fixed_init_state:
self._init_state = self.observation_space.sample()
self._done = False
self._reward = 0.0
self._cum_reward = 0.0
self._info = {'actions': []}
self._action = self.simulator.get_first_action()
self._actions = []
self._first_step = True
self._step = 0
obs = np.array(self.simulator.reset(self._init_state))
# if self.blackbox_sim_state:
# obs = np.array([0] * self.action_space.shape[0])
# else:
# print('Not action only')
if not self.blackbox_sim_state:
obs = np.concatenate((obs, np.array(self._init_state)), axis=0)
# self.root_action = self.downsample(self._action)
self.root_action = self._action
# obs = np.concatenate((np.array([self._step]), self.downsample(obs)), axis=0)
return obs
@property
def action_space(self):
"""
Returns a Space object
"""
if self.spaces is None:
# return self._to_garage_space(self.simulator.action_space)
return self.simulator.action_space
else:
return self.spaces.action_space
@property
def observation_space(self):
"""
Returns a Space object
"""
if self.spaces is None:
# return self._to_garage_space(self.simulator.observation_space)
return self.simulator.observation_space
else:
return self.spaces.observation_space
def get_cache_list(self):
return self._info
def log(self):
self.simulator.log()
def render(self, **kwargs):
if hasattr(self.simulator, "render") and callable(
getattr(self.simulator, "render")):
return self.simulator.render(**kwargs)
else:
return None
def close(self):
if hasattr(self.simulator, "close") and callable(
getattr(self.simulator, "close")):
self.simulator.close()
else:
return None
def vec_env_executor(self, n_envs, max_path_length):
return self.simulator.vec_env_executor(n_envs, max_path_length,
self.reward_function,
self._fixed_init_state,
self._init_state,
self.open_loop)
def log_diagnostics(self, paths):
pass
@cached_property
def spec(self):
"""
Returns an EnvSpec.
Returns:
spec (garage.envs.EnvSpec)
"""
return EnvSpec(observation_space=self.observation_space,
action_space=self.action_space)
def get_params_internal(self, **tags):
# this lasagne function also returns all var below the passed layers
if not self.params_set:
self.p_db_filename = GoExploreParameter("db_filename",
self.db_filename)
self.p_key_list = GoExploreParameter("key_list", self.key_list)
self.p_max_value = GoExploreParameter("max_value", self.max_value)
self.p_robustify_state = GoExploreParameter(
"robustify_state", self.robustify_state)
self.params_set = True
if tags.pop("db_filename", False):
return [self.p_db_filename]
if tags.pop("key_list", False):
return [self.p_key_list]
if tags.pop("max_value", False):
return [self.p_max_value]
if tags.pop("robustify_state", False):
return [self.p_robustify_state]
return [
self.p_db_filename, self.p_key_list, self.p_max_value,
self.p_robustify_state
] # , self.p_downsampler]
def set_param_values(self, param_values, **tags):
debug = tags.pop("debug", False)
for param, value in zip(self.get_params(**tags), param_values):
param.set_value(value)
if debug:
print("setting value of %s" % param.name)
def get_param_values(self, **tags):
return [
param.get_value(borrow=True) for param in self.get_params(**tags)
]
def downsample(self, obs):
return obs
def _get_obs(self):
return self.simulator._get_obs()
def run_task(snapshot_config, *_):
seed = 0
# top_k = 10
np.random.seed(seed)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
with tf.variable_scope('AST', reuse=tf.AUTO_REUSE):
with LocalTFRunner(snapshot_config=snapshot_config,
max_cpus=4,
sess=sess) as local_runner:
# Instantiate the example classes
sim = ExampleAVSimulator(**sim_args)
reward_function = ExampleAVReward(**reward_args)
spaces = ExampleAVSpaces(**spaces_args)
# Create the environment
if 'id' in env_args:
env_args.pop('id')
env = ASTEnv(simulator=sim,
reward_function=reward_function,
spaces=spaces,
**env_args)
env = TfEnv(env)
policy = ContinuousMLPPolicy(name='ast_agent',
env_spec=env.spec,
**policy_args)
params = policy.get_params()
sess.run(tf.variables_initializer(params))
# Instantiate the garage objects
baseline = ZeroBaseline(env_spec=env.spec)
top_paths = BPQ.BoundedPriorityQueue(**bpq_args)
sampler_cls = ASTVectorizedSampler
sampler_args = {
"open_loop": False,
"sim": sim,
"reward_function": reward_function,
"n_envs": n_parallel
}
if ga_type == 'ga':
print('ga')
algo = GA(env_spec=env.spec,
policy=policy,
baseline=baseline,
top_paths=top_paths,
**algo_args)
elif ga_type == 'gasm':
print('gasm')
algo = GASM(env_spec=env.spec,
policy=policy,
baseline=baseline,
top_paths=top_paths,
**algo_args)
else:
raise NotImplementedError
local_runner.setup(algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args=sampler_args)
# Run the experiment
local_runner.train(**runner_args)
def test_example_av_simulator():
simulator_args = {'car_init_x': 0,
'car_init_y': 0}
sim = ExampleAVSimulator(simulator_args=simulator_args, max_path_length=1)
sim.blackbox_sim_state = False
# check reset
init = sim.reset(s_0=np.array([0, 0, 0, 0, 0]))
assert np.all(init == np.array([0, 0, 0, 0]))
# Check if simulate can find a goal
path_length, info = sim.simulate(actions=[np.zeros(6)], s_0=np.array([0, 0, 0, 1, 0]))
assert path_length == 0
# Check if simulate can return end of path
path_length, info = sim.simulate(actions=[np.zeros(6)], s_0=np.array([5, 5, 0, 0, 0]))
assert path_length == -1
# Check open-loop sim step
sim.open_loop = True
sim.reset(s_0=np.array([0, 0, 0, 0, 0]))
obs = sim.step(action=np.zeros(6))
assert np.all(obs == np.array([0, 0, 0, 0, 0]))
# Check closed-loop sim step
sim.open_loop = False
init = sim.reset(s_0=np.array([0, 0, 0, 0, 0]))
obs = sim.step(action=np.zeros(6))
assert np.all(obs == np.array([0, 0, 0, 0]))
def __init__(self,
open_loop=True,
blackbox_sim_state=True,
fixed_init_state=False,
s_0=None,
simulator=None,
reward_function=None,
spaces=None):
# gym_env = gym.make('ast_toolbox:GoExploreAST-v0', {'test':'test string'})
# pdb.set_trace()
# super().__init__(gym_env)
# Constant hyper-params -- set by user
self.open_loop = open_loop
self.blackbox_sim_state = blackbox_sim_state # is this redundant?
self.spaces = spaces
if spaces is None:
self.spaces = ExampleAVSpaces()
# These are set by reset, not the user
self._done = False
self._reward = 0.0
self._info = {}
self._step = 0
self._action = None
self._actions = []
self._first_step = True
self.reward_range = (-float('inf'), float('inf'))
self.metadata = None
self.spec._entry_point = []
self._cum_reward = 0.0
self.root_action = None
self.sample_limit = 10000
self.simulator = simulator
if self.simulator is None:
self.simulator = ExampleAVSimulator()
if s_0 is None:
self._init_state = self.observation_space.sample()
else:
self._init_state = s_0
self._fixed_init_state = fixed_init_state
self.reward_function = reward_function
if self.reward_function is None:
self.reward_function = ExampleAVReward()
if hasattr(self.simulator, "vec_env_executor") and callable(getattr(self.simulator, "vec_env_executor")):
self.vectorized = True
else:
self.vectorized = False
# super().__init__(self)
# Always call Serializable constructor last
self.params_set = False
self.db_filename = 'database.dat'
self.key_list = []
self.max_value = 0
self.robustify_state = []
self.robustify = False
Parameterized.__init__(self)
