replay-and-data

# Replay and Data Use this sub-skill when an Acme task involves collecting actor experience, writing it to Reverb, reading replay samples into a learner, adapting offline demonstrations, or diagnosing replay/data structure mismatches. ## Read First - [Replay API](references/replay-api.md): `Adder` contracts, Reverb adder families, signatures, dataset helpers, and expected shapes. - [Replay workflows](references/workflows.md): choose transition/sequence/episode/structured adders, build Reverb/TFDS/numpy iterators, and connect learner data. - [Troubleshooting](references/troubleshooting.md): optional dependency failures, Reverb table signature errors, sequence/period pitfalls, extras mismatches, and invalid offline data paths. - [Replay structure helper](scripts/describe_replay_structure.py): validate a JSON description of observation/action/reward/discount/extras nesting and print an adder-family recommendation without importing Reverb. ## Route Boundaries - Stay here for `acme.adders.base`, `acme.adders.wrappers`, `acme.adders.reverb.*`, `acme.datasets.reverb`, `acme.datasets.tfds`, `acme.datasets.numpy_iterator`, `acme.datasets.image_augmentation`, and `acme.utils.reverb_utils`. - Use the sibling `core-workflows` sub-skill for environment-loop basics, wrapper selection, `dm_env` conversion, experiment scaffolding, or actor/learner process layout before replay details matter. - Use the sibling `jax-agents` or `tf-agents` sub-skill when the question is which algorithm-specific builder, config, or learner expects a particular replay table or iterator shape. ## Fast Choices - Pick `NStepTransitionAdder` for feed-forward learners that train on `(observation, action, reward, discount, next_observation, extras)` transitions. - Pick `SequenceAdder` for recurrent learners, sequence losses, overlapping unrolls, or learner code that expects time-major/batched trajectories. - Pick `EpisodeAdder` for full-episode imitation/offline pipelines or algorithms that need complete trajectories. - Pick `StructuredAdder` when one actor stream must populate multiple tables or custom item patterns with Reverb `structured_writer` configs. - For offline demonstrations, adapt data into `types.Transition` batches or Reverb-compatible samples, then expose an iterator matching the learner constructor.