`evorl.algorithms.erl.erl_workflow`¶

Module Contents¶

Classes¶

`ERLTrainMetric`
`ERLWorkflowBase`
`WorkflowMetric`

API¶

class evorl.algorithms.erl.erl_workflow.ERLTrainMetric[source]¶

Bases: evorl.metrics.MetricBase

ec_info: evorl.types.PyTreeDict¶: ‘metric_field(…)’

pop_episode_lengths: chex.Array | None¶: None

pop_episode_returns: chex.Array | None¶: None

rb_size: chex.Array | None¶: None

rl_episode_lengths: chex.Array | None¶: None

rl_episode_returns: chex.Array | None¶: None

rl_metrics: evorl.metrics.MetricBase | None¶: None

class evorl.algorithms.erl.erl_workflow.ERLWorkflowBase(*, env: evorl.envs.Env, agent: evorl.agent.Agent, agent_state_vmap_axes: evorl.agent.AgentStateAxis, optimizer: optax.GradientTransformation, ec_optimizer: evorl.ec.optimizers.EvoOptimizer, ec_collector: evorl.evaluators.EpisodeCollector, rl_collector: evorl.evaluators.EpisodeCollector, evaluator: evorl.evaluators.Evaluator, replay_buffer: evorl.replay_buffers.AbstractReplayBuffer, config: omegaconf.DictConfig)[source]¶

Bases: evorl.workflows.Workflow

classmethod build_from_config(config: omegaconf.DictConfig, enable_multi_devices: bool = False, enable_jit: bool = True) → typing_extensions.Self[source]¶

classmethod enable_jit() → None[source]¶

abstract evaluate(state: evorl.types.State) → tuple[evorl.metrics.MetricBase, evorl.types.State][source]¶

setup(key: chex.PRNGKey) → evorl.types.State[source]¶

abstract warmup_step(state: evorl.types.State) → tuple[evorl.metrics.MetricBase, evorl.types.State][source]¶

class evorl.algorithms.erl.erl_workflow.WorkflowMetric[source]¶

Bases: evorl.metrics.MetricBase

iterations: chex.Array¶: ‘zeros(…)’

rl_sampled_timesteps: chex.Array¶: ‘zeros(…)’

sampled_episodes: chex.Array¶: ‘zeros(…)’

sampled_timesteps: chex.Array¶: ‘zeros(…)’