def default_config():
"""Default configuration values."""
data_root = serialize.get_output_dir() # where models are read from
log_root = serialize.get_output_dir() # where results are written to
n_bootstrap = 1000 # number of bootstrap samples
alpha = 95 # percentile confidence interval
aggregate_kinds = ("bootstrap", "studentt", "sample")
_ = locals()
del _
def default_config():
"""Default configuration values."""
data_root = serialize.get_output_dir() # where values are read from
log_root = serialize.get_output_dir() # where results are written to
vals_path = None # aggregated data values
# Reward configurations: models to compare
x_reward_cfgs = None
y_reward_cfgs = None
heatmap_kwargs = {}
_ = locals()
del _
def default_config():
"""Default configuration values."""
# Reward parameters
exp_name = "default"
discount = 0.99
# Figure parameters
vmin = None
vmax = None
log_root = os.path.join(serialize.get_output_dir(), "plot_gridworld_reward")
fmt = "pdf" # file type
styles = ["paper", "gridworld-heatmap", "gridworld-heatmap-2in1"]
ncols = 2
# We can't use actual LaTeX since it draws too thick hatched lines :(
# This means we can't use our macros `figsymbols.sty`, so manually transcribe them
# to be as close as possible using matplotlib native features.
rewards = [
(r"$\mathtt{Sparse}$", "sparse_goal"),
(r"$\mathtt{Dense}$", "transformed_goal"),
(r"$\mathtt{Penalty}$", "sparse_penalty"),
(r"$\mathtt{Center}$", "center_goal"),
(r"$\mathtt{Path}$", "dirt_path"),
(r"$\mathtt{Cliff}$", "cliff_walk"),
]
_ = locals() # quieten flake8 unused variable warning
del _
def default_config():
"""Default configuration."""
log_root = serialize.get_output_dir() # where results are written to
configs = {}
run_tag = "default"
_ = locals()
del _
def _canonicalize_data_root(path: str) -> str:
if path.endswith("dummy"):
path = "dummy"
for root_prefix in DATA_ROOT_PREFIXES:
if path.startswith(root_prefix):
path = path.replace(root_prefix, serialize.get_output_dir())
break
return path
def default_config():
"""Default configuration values."""
env_name = "evaluating_rewards/PointMassLine-v0"
kinds = POINT_MASS_KINDS
data_root = serialize.get_output_dir() # where models are read from
log_root = serialize.get_output_dir() # where results are written to
n_bootstrap = 1000 # number of bootstrap samples
alpha = 95 # percentile confidence interval
aggregate_kinds = ("bootstrap", "studentt", "sample")
vals_path = None
# Reward configurations: models to compare
x_reward_cfgs = None
y_reward_cfgs = None
_ = locals()
del _
def point_maze():
"""IRL config for PointMaze environment."""
env_name = "imitation/PointMazeLeftVel-v0"
rollout_path = os.path.join(
serialize.get_output_dir(),
"train_experts/ground_truth/20201203_105631_297835/imitation_PointMazeLeftVel-v0",
"evaluating_rewards_PointMazeGroundTruthWithCtrl-v0/best/rollouts/final.pkl",
)
total_timesteps = 1e6
_ = locals()
del _
def experiment_main(experiment: sacred.Experiment,
name: str,
sacred_symlink: bool = True):
"""Returns a main function for experiment."""
sacred_dir = os.path.join(serialize.get_output_dir(), "sacred", name)
observer = observers.FileStorageObserver.create(sacred_dir)
if sacred_symlink:
experiment.pre_run_hook(add_sacred_symlink(observer))
experiment.observers.append(observer)
experiment.run_commandline()
def point_maze_learned_checkpoint_cfgs(
prefix: str = "transfer_point_maze",
target_num: Optional[int] = None,
shard: Optional[Tuple[int, int]] = None,
) -> List[RewardCfg]:
"""Configurations for learned rewards in PointMaze for each checkpoint.
Args:
prefix: The directory to locate results under.
target_num: The target number of checkpoints to return for each algorithm.
May return less than this if there are fewer checkpoints than `target_num`.
May return up to twice this number.
shard: Optional tuple of `(shard_num, total_shards)`. If specified, partitions
checkpoints into `total_shards` of (approximately) equally-spaced checkpoints,
returning only the `shard_num`'th partition. This gives a simple way to parallelize,
or just run sequentially with increasing levels of resolution but still full coverage.
Returns:
Reward configurations for checkpoints for each reward model.
"""
res = {}
for kind, fstr_path in _POINT_MAZE_CFG:
glob_path = os.path.join(
serialize.get_output_dir(), prefix, "reward", fstr_path.format("[0-9]*")
)
paths = sorted(glob.glob(glob_path))
cfgs = [(kind, path) for path in paths]
if target_num and len(cfgs) > target_num:
subsample = math.floor(len(cfgs) / target_num)
cfgs = cfgs[::subsample]
assert target_num <= len(cfgs) <= 2 * target_num
if shard:
shard_num, total_shards = shard
cfgs = cfgs[shard_num::total_shards]
res[(kind, fstr_path)] = cfgs
empty = {k: bool(v) for k, v in res.items()}
if any(empty) and not all(empty):
raise ValueError(f"No checkpoints found for some algorithms: {empty}")
return list(itertools.chain(*res.values()))
def default_config():
"""Default configuration for combined_distances."""
vals_paths = []
log_root = serialize.get_output_dir(
) # where results are read from/written to
experiment_kinds = {}
distance_kinds_order = ("epic", "npec", "erc", "rl")
config_updates = {} # config updates applied to all subcommands
named_configs = {}
skip = {}
target_reward_type = None
target_reward_path = None
pretty_models = {}
pretty_algorithms = {}
# Output formats
output_fn = latex_table
styles = ["paper", "tex", "training-curve", "training-curve-1col"]
tag = "default"
_ = locals()
del _
def default_config():
"""Default configuration values."""
normalize = False
# Dataset parameters
log_root = serialize.get_output_dir(
) # where results are read from/written to
discount = 0.99
reward_subset = None
# Figure parameters
kind = "npec"
styles = [
"paper", "heatmap", "heatmap-1col", "heatmap-1col-fatlabels", "tex"
]
save_kwargs = {
"fmt": "pdf",
}
_ = locals()
del _
def add_logging_config(experiment, name):
experiment.add_config(
{"log_root": os.path.join(serialize.get_output_dir(), name)})
experiment.config(logging_config)
def add_logging_config(experiment: sacred.Experiment, name: str) -> None:
experiment.add_config(
{"log_root": os.path.join(serialize.get_output_dir(), name)})
experiment.config(logging_config)
