module moplayground.config

Single access point between YAML config files and the moplayground package.

This is the only module that knows YAML key names. Scripts call :func:load to read a config, then use the getter functions to pull plain values out of it and pass those values to package functions. Package code never indexes a config directly.

Canonical layout:

     name, save_dir, description, algorithm, env, backend, gaitlib_path      env_config: {...}      learning_params:        ppo_params:           {learning_rate, num_envs, ...}        sampling_params:      {alpha, k, sampling, warmup_frac}        morlax_params:        {hypertype, hypersize, num_features,                               policy_hidden_layer_sizes, value_hidden_layer_sizes}        amor_params:          {policy_hidden_layer_sizes, value_hidden_layer_sizes}        morlax_warmup_params: {enabled, policy} 

:func:load renames keys from older layouts to the canonical layout, so saved config.yaml files from old runs still load.

Global Variables

• ALGORITHMS
• ENVS
• SAMPLING_MODES
• HYPERTYPES

---

function load

load(path) → ConfigDict

Read a YAML config, convert it to the canonical layout, and validate it.

Args:

• path: Path to a YAML config file. Either a file in config/ or a config.yaml saved in a run directory.

Returns: ConfigDict in the canonical layout.

Raises:

• ValueError: If the config fails validation. The message lists every problem found.

---

function save

save(cfg, run_dir, warn_github_changes=True) → Path

Create run_dir and write config.yaml into it.

A run named test may reuse an existing run_dir and does not record a git hash. Any other name raises if run_dir already exists, and records the current git hash as cfg.git_hash.

Args:

• cfg: Config returned by :func:load. Modified in place (git_hash).
• run_dir: Run directory, normally :func:run_dir of cfg.
• warn_github_changes: Forwarded to the git-hash lookup.

Returns: Path to the written config.yaml.

---

function algorithm

algorithm(cfg) → str

'morlax' or 'amor'.

---

function env_name

env_name(cfg) → str

Environment class name, e.g. 'MOCheetah'.

---

function backend

backend(cfg, for_training=False) → str

'jnp' when building an env for training, else the configured backend.

---

function env_params

env_params(cfg) → ConfigDict

The env_config section, given to env classes as env_params.

---

function gaitlib_path

gaitlib_path(cfg)

Gait library path (NaviGait only), or None.

---

function ppo_params

ppo_params(cfg) → dict

PPO settings shared by both algorithms.

---

function sampling_params

sampling_params(cfg) → dict

Preference-sampling settings: alpha, k, sampling, warmup_frac.

---

function network_params

network_params(cfg) → dict

Network settings for the configured algorithm.

---

function normalize_observations

normalize_observations(cfg) → bool

Whether training keeps a running observation normalizer.

---

function objectives

objectives(cfg) → list

Reward keys per objective, one list per reward dimension.

---

function num_objectives

num_objectives(cfg) → int

Number of objectives (length of the reward vector).

---

function objective_labels

objective_labels(cfg)

Display names for the objectives, or None if not set.

---

function run_name

run_name(cfg) → str

Run name.

---

function run_dir

run_dir(cfg) → Path

Run directory: save_dir / name.

---

function check_same_env

check_same_env(cfgs)

Raise ValueError unless every config deploys the same environment.

Compares env and the full env_config section. Use this before evaluating several runs together (e.g. eval.pareto.get_morlax_fronts with a list of run directories).

---

function set_name

set_name(cfg, value)

---

function set_hypertype

set_hypertype(cfg, value)

---

function set_sampling

set_sampling(cfg, value)

---

function set_k

set_k(cfg, value)

---

function apply_overrides

apply_overrides(cfg, overrides, where='overrides')

Write each value in a nested dict into cfg at the same key path.

overrides mirrors the config layout, e.g. {'learning_params': {'ppo_params': {'learning_rate': 3e-4}}}. A wandb sweep with nested parameters gives its sampled values in this form.

Args:

• cfg: Config returned by :func:load. Modified in place.
• overrides: Nested dict of values to write.
• where: Name for overrides in error messages.

Raises:

• ValueError: If a key path in overrides does not exist in cfg (e.g. a typo in a sweep file), or if the result fails validation.