def logging_config(log_root, search):
log_dir = os.path.join( # noqa: F841 pylint:disable=unused-variable
log_root,
"plot_epic_heatmap",
str(search).replace("/", "_"),
util.make_unique_timestamp(),
)
def logging_config(log_root, tag):
"""Default logging configuration: hierarchical directory structure based on config."""
log_dir = os.path.join( # noqa: F841 pylint:disable=unused-variable
log_root,
"combined_distances",
tag,
imit_util.make_unique_timestamp(),
)
def logging_config(log_root, env_name, dataset_tag, corr_kind, discount):
"""Default logging configuration: hierarchical directory structure based on config."""
log_dir = os.path.join( # noqa: F841 pylint:disable=unused-variable
log_root,
"erc",
env_name,
dataset_tag,
corr_kind,
f"discount{discount}",
imit_util.make_unique_timestamp(),
)
def paths(env_name, log_root, rollout_hint, data_dir):
log_dir = os.path.join(log_root, env_name.replace("/", "_"),
util.make_unique_timestamp())
# Recommended that user sets rollout_path manually.
# By default we guess the named config associated with `env_name`
# and attempt to load rollouts from `data/expert_models/`.
if rollout_hint is None:
rollout_hint = env_name.split("-")[0].lower()
rollout_path = os.path.join(data_dir, "expert_models", f"{rollout_hint}_0",
"rollouts", "final.pkl")
def logging_config(env_name, log_root):
"""Default logging configuration: hierarchical directory structure based on config."""
timestamp = imit_util.make_unique_timestamp()
log_dir = os.path.join(
log_root,
"plot_heatmap",
env_name,
timestamp,
)
_ = locals()
del _
def hook(config, command_name, logger):
del logger
updates = {}
if config["common"]["log_dir"] is None:
env_sanitized = config["common"]["env_name"].replace("/", "_")
log_dir = os.path.join(
"output",
command_name,
env_sanitized,
util.make_unique_timestamp(),
)
updates["log_dir"] = log_dir
return updates
def step(self,
user_action: np.ndarray) -> Tuple[np.ndarray, float, bool, dict]:
"""Steps the environment.
DAgger needs to be able to inject imitation policy actions randomly at some
subset of time steps. This method will replace the given action with a
"robot" (i.e. imitation policy) action if necessary.
Args:
user_action: the _intended_ demonstrator action for the current
state. This will be executed with probability `self.beta`.
Otherwise, a "robot" action will be sampled and executed
instead.
Returns:
next_obs, reward, done, info: unchanged output of `self.env.step()`.
"""
assert self._is_reset, "call .reset() before .step()"
# Replace the given action with a robot action 100*(1-beta)% of the time.
if np.random.uniform(0, 1) > self.beta:
actual_act = self.get_robot_act(self._last_obs)
else:
actual_act = user_action
# actually step the env & record data as appropriate
next_obs, reward, done, info = self.env.step(actual_act)
self._last_obs = next_obs
self.traj_accum.add_step({
"acts": user_action,
"obs": next_obs,
"rews": reward,
"infos": info
})
# if we're finished, then save the trajectory & print a message
if done and not self._done_before:
trajectory = self.traj_accum.finish_trajectory()
timestamp = util.make_unique_timestamp()
trajectory_path = os.path.join(self.save_dir,
"dagger-demo-" + timestamp + ".npz")
logging.info(f"Saving demo at '{trajectory_path}'")
_save_trajectory(trajectory_path, trajectory)
if done:
# record the fact that we're already done to avoid saving demo over and
# over until the user resets
self._done_before = True
return next_obs, reward, done, info
def config():
sacred_ex_name = "expert_demos" # The experiment to parallelize
_uuid = make_unique_timestamp()
run_name = (f"DEFAULT_{_uuid}"
) # CLI --name option. For analysis grouping.
resources_per_trial = {} # Argument to `tune.run`
base_named_configs = [
] # Background settings before search_space is applied
base_config_updates = {
} # Background settings before search_space is applied
search_space = {
"named_configs": [],
"config_updates": {},
} # `config` argument to `ray.tune.run(trainable, config)`
local_dir = None # `local_dir` arg for `ray.tune.run`
upload_dir = None # `upload_dir` arg for `ray.tune.run`
n_seeds = 3 # Number of seeds to search over by default
def logging_config(log_root):
log_dir = os.path.join( # noqa: F841 pylint:disable=unused-variable
log_root,
"plot_gridworld_divergence",
util.make_unique_timestamp(),
)
def logging(env_name, log_root):
log_dir = os.path.join(log_root, env_name.replace("/", "_"),
util.make_unique_timestamp())
def logging_config(log_root, env_name):
log_dir = os.path.join(log_root, env_name.replace("/", "_"),
util.make_unique_timestamp())
_ = locals() # quieten flake8 unused variable warning
del _
def logging_config(log_root, exp_name):
log_dir = os.path.join( # noqa: F841 pylint:disable=unused-variable
log_root,
exp_name,
util.make_unique_timestamp(),
)
def logging_config(log_root, run_tag):
"""Logging configuration: timestamp plus unique UUID."""
log_dir = os.path.join(log_root, "train_experts", run_tag,
util.make_unique_timestamp())
_ = locals()
del _
