unilab.algos.torch.offpolicy.double_buffer_runner.DoubleBufferOffPolicyRunner

class unilab.algos.torch.offpolicy.double_buffer_runner.DoubleBufferOffPolicyRunner[source]

Bases: OffPolicyRunner

OffPolicyRunner variant that uses CPUPinnedDoubleBufferReplayPipeline.

The only behavioural difference from the parent class is in learn(): - ReplayBuffer is created as packed CPU shared storage. - Sampling goes through CPUPinnedDoubleBufferReplayPipeline instead of

ReplayBuffer.sample().

Parameters:

  • replay_prefetch_mode (str)

  • verbose_metrics (bool)

Methods

__init__(*[, replay_prefetch_mode, ...])

close()

learn([max_iterations, save_interval, ...])

Unified training loop for off-policy algorithms.

Attributes

LEARNER_LOG_INTERVAL

LEARNER_LOG_INTERVAL = 10
__init__(*, replay_prefetch_mode='one_tick', verbose_metrics=False, **kwargs)[source]
Parameters:

  • replay_prefetch_mode (str)

  • verbose_metrics (bool)

replay_transfer_backend: dict[str, object]
learn(max_iterations=1500, save_interval=50, log_dir='logs', logger_type='tensorboard')[source]

Unified training loop for off-policy algorithms.

Parameters:

  • max_iterations (int)

  • save_interval (int)

  • log_dir (str)

  • logger_type (str)

Return type:

None

close()
Return type:

None