aac_datasets.utils.collate module

class AdvancedCollate(

fill_values: Dict[str, float | int],

)

Bases: object

Advanced collate object for DataLoader.

Merge lists in dicts into a single dict of lists. Audio will be padded if a fill value is given in __init__.

Example

>>> collate = AdvancedCollate({"audio": 0.0})
>>> loader = DataLoader(..., collate_fn=collate)
>>> next(iter(loader))
... {"audio": tensor([[...]]), ...}

class BasicCollate

Bases: object

Collate object for DataLoader.

Merge lists in dicts into a single dict of lists. No padding is applied.

can_be_stacked(

tensors: List[Tensor],

) → bool

Returns true if a list of tensors can be stacked with torch.stack function.

pad_last_dim(

tensor: Tensor,

target_length: int,

pad_value: float,

) → Tensor

Left padding tensor at last dim.

Parameters:

• tensor – Tensor of at least 1 dim. (…, T)

• target_length – Target length of the last dim. If target_length <= T, the function has no effect.

• pad_value – Fill value used to pad tensor.

Returns:

A tensor of shape (…, target_length).