Return attribute value if present, else default (dict-like API). Args: key: Attribute name. default: Fallbac

Append value to list (only for ListConfig)

Return config keys (required for dict unpacking) Tracks all keys as accessed. Only works for DictConfig.

Iterate (key, value) pairs like dict.items(). Returns: Generator[Tuple[str, Any], None, None]

Remove and return a value

Return a shallow copy (does not copy access tracking state)

Get value with default fallback

Ensure state/action concat is always appended as the last transform.

Update config with values from another dict/config

Return config values (tracks all keys as accessed)

Convert tensor back to numpy array efficiently. Args: frames: tensor of shape [T, C, H, W] in range [0, 1] Returns:

Set default value if key doesn't exist

Extend list with values (only for ListConfig)

Check if underlying config is a ListConfig

Restore a model instance from a saved checkpoint. Workflow: 1. Resolve checkpoint path 2. Load config + data

Return config items (tracks all keys as accessed)

Return a deep copy (does not copy access tracking state)

embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) returns: (M, D)

Load framework config and dataset normalization statistics for a saved checkpoint. Args: pretrained_checkpoint: Path to a .pt checkp

Combine statistics from all sub-keys under a modality. Args: modality_stats (dict): Statistics for a modality, containing multip

Set the epoch for the dataset. Args: epoch (int): The epoch to set.

Get the underlying OmegaConf object

Limit decoder threading to reduce host-memory spikes under many DataLoader workers.

Cache episode metadata to avoid repeated lookups on hot paths.

Summarize tensor scale with norm/value quantiles for spike debugging.

Get root config object

Extract used action and state keys from modality configuration.

Set the metadata for the transforms. This is useful for transforms that need to know the metadata, such as the normalization values.

execute training loop

Seed EMA statistics from current codebooks.